# AI Spatial Audio: From Research to Creative Control

> Spatial9 reviews 124 AI spatial audio papers, from source separation and binaural rendering to room acoustics and codecs, with 15 code examples and verdicts.

Web page: https://spatial9.ai/blog/ai-spatial-audio-research-review
Author: Spatial9 Engineering Team
Published: 2026-10-25. Updated: 2026-10-25.

## Key takeaways
- Spatial9 reviewed 124 academic papers on AI spatial audio across twelve research branches, from source separation and localization to neural acoustics, HRTF personalization, sound zones, codecs and evaluation.
- The central position: creative intent must remain explicit, editable and testable from source to playback.
- Separation quality, inventory completeness and persistent source identity are three separate judgments; a broad stem is not always the object a creator needs to move.
- Event timing and trajectory design are distinct responsibilities: SEND decides when an object should move, while a separate path defines how it moves.
- Room models, HRTF profiles and sound-zone controllers carry operating contracts such as calibration, complete filters with timing, and measured acoustic paths.
- Spatial quality is a profile of distinct attributes (fidelity, localization, externalization, immersion, preference), so no single metric score is enough.
- Fifteen code examples connect the research to inspectable engineering decisions, including official HTDemucs and SAM Audio API usage.

Short answer: AI spatial audio is no longer one technology but a set of specialized capabilities: source separation, localization, binaural and Ambisonic rendering, neural room acoustics, HRTF personalization, sound-zone control, neural codecs and quality metrics. After reviewing 124 academic papers, the Spatial9 Engineering Team concludes that production value depends less on any single model and more on keeping creative intent explicit, editable and testable from the first source file to final playback.

This paper presents Spatial9's technical assessment of artificial intelligence for spatial and immersive audio. It discusses all ninety-six academic records in the original research catalog and twenty-eight additional records covering spatial reasoning, binaural enhancement, loudspeaker control, production interaction and differentiable processing. Every record receives an individual account of its research contribution, a Spatial9 assessment and a direct verdict on its role in production. Standards and company descriptions remain distinct from academic results. Fifteen short code examples connect selected papers to explicit implementation decisions, from source identity and movement timing to acoustic control and spatial fidelity.

Our central position is that production value depends on preserving creative intent across a chain of decisions. Source fidelity, stable object identity, intentional motion, acoustic plausibility, listener variation, revision effort and delivery behavior all contribute to that value. We judge each method on its stated task, then assess the relationship between that contribution and a creator's ability to control a complete session. A specialist component can be important without supplying every part of the workflow.

This is a bounded critical review of 124 academic records, with access depth and reviewed versions disclosed. Findings attributed to the research are the original authors' reported findings; the assessments and verdicts express Spatial9's interpretation of their production relevance. The resulting position is clear: creative intent needs an explicit representation, meaningful controls and an audible consequence that remains coherent through revision and playback. We apply that position to the literature and to our own architectural priorities.

This article publishes that review in full: every one of the 124 paper assessments, all fifteen code examples with their reading notes and scope limits, the research task map, our fourteen assessment criteria and the principles that follow from them. Each paper review has three parts: the research contribution as its authors report it, the Spatial9 assessment and our verdict on its role in production.

[Image: The twelve research branches covered by this review, connected by one production question: does the creator keep control of the decision?]


## How to read this review

This manuscript provides an individually traceable discussion of the complete original research catalog, addresses important omissions found during the coverage audit and states Spatial9's position across the field. The coverage index links every identifier to a dedicated review. The catalog includes models, datasets, metrics, surveys, codecs and enabling methods. Each has a defined role; inclusion does not imply that every record is a direct product competitor.

Coverage is complete relative to the declared 124-record catalog. The wider literature spans speech processing, computer vision, acoustics, hearing science, graphics, robotics, music technology and human-

computer interaction. The review makes its boundaries visible through the source list, version metadata and access descriptions. Those boundaries support a precise discussion of the included research without claiming a census of every publication worldwide.

Each review has three parts. Research contribution describes the source and the scope of its evidence. Spatial9 assessment explains our technical interpretation, including its strengths and limits for immersive production. Our verdict states the method's practical role and the resulting product or architecture implication. A narrow task can be a valuable specialization; we distinguish that specialization from a capability the authors never claim to supply.

Readers seeking our company position can begin with the executive argument, the cross-paper synthesis and the assessment criteria. Technical readers can follow the task map and paper chapters. Editors can use the publication sequence to develop focused articles with the relevant research context. Primary links and the coverage index make the judgments traceable to the original records.


### The Spatial9 position

Our organizing question is whether an AI system helps a person make, revise and deliver a spatial decision. A binaural generator, separator, scene representation, acoustic field and motion planner contribute different capabilities. We see product value in making those capabilities cooperate while keeping the creator's intention recoverable when the brief changes.

Spatial9's position is that creative intent must remain explicit, editable and testable from source to playback. This connects the first result with the decisions that follow: a revised arrangement, a repeated chorus, a moving listener and a file opened by the recipient. Our technical assessments prioritize that complete relationship. A successful component receives credit for its contribution, with the remaining production responsibilities identified clearly.

Important building blocks already exist in the literature. STASE connects semantic planning with spatial rendering. Differentiable mixing exposes adjustable parameters. SEND examines intentional movement timing, and S3A provides a history of integrated object-audio workflows. We credit that lineage and locate Spatial9's architectural priorities in the continuity between analysis, scene control, revision and delivery. [7, 8, 1, 3, 5]

We judge methods according to their input contract and intended use. Speech enhancement, binaural reproduction, horizontal Ambisonics and music upmixing address different questions. Two output channels can support a dynamic binaural experience, and a deliberately horizontal representation can be appropriate for its task. Our assessments identify the contribution first and then explain its suitability for the production setting under discussion.

We distinguish a command from its audible consequence. A saved coordinate describes an intended position; the rendered signal and listener experience determine its perceptual effect. Internal object controls and exported object structures describe different stages. Throughput and interactive delay describe different operating properties. Our judgments retain those distinctions because they affect what a creator can reasonably expect.

We consider the complete session the unit of production usefulness. Source identity, repeated sections, silence, tails, chunk boundaries and corrective edits all contribute to coherence. Dynamic and streaming capabilities are credited where the reviewed work supplies them. Our position adds a consistent question across methods: how does the capability remain useful as the scene and the creator's decisions evolve?

We keep the dimensions of quality visible. Fidelity, localization, externalization, immersion and artistic preference are related but distinct. SAQAM, AMBIQUAL, QASTAnet, spatial impairment decomposition and perceptual-loss research provide instruments with different domains. Our assessment favors an interpretable profile of outcomes over a single unexplained score, especially when improvements in one attribute coincide with a tradeoff in another. [116, 117, 118, 119, 120, 124]

Our critical voice is technically generous and demanding about the meaning of a claim. We recognize valuable specialized methods, identify the assumptions that affect their use and explain the product decisions those assumptions imply. The same criteria apply to Spatial9. Our authority rests on a clear connection between a technical contribution, its evidence and the creator's practical control of the resulting experience.


### Review method and evidence boundaries

The starting set consists of ninety-six academic records assembled in the preceding research roadmap: twenty-five binaural/upmixing records, twenty-eight separation/understanding/generation records, twenty-

seven acoustics/HRTF records and sixteen planning/evaluation/representation records. This expansion preserves every original identifier and adds twenty-eight records. The final total is 124 academic records, plus four standards or assessment recommendations. Related versions and extensions are labeled so that paper count is not confused with the number of independent model families.

Discovery used task-specific searches and reference following across IEEE/ICASSP/TASLP, AES, DAFx, ISCA/Interspeech, ACM, CVPR/ICCV/ECCV, NeurIPS, ICML, ICLR, ISMIR, DCASE, ACL-family proceedings, arXiv and author or institutional repositories. Primary material was preferred: official proceedings, author manuscripts, project pages and author-maintained repositories. Secondary search results were used for discovery, with substantive technical claims checked against primary sources.

Search families included mono-to-binaural conversion; music upmixing; Ambisonic generation and codecs; promptable and music source separation; sound-event localization and tracking; spatial audio-

language reasoning; audio-visual and 360-degree synthesis; neural acoustic fields and room impulse responses; HRTF personalization; binaural enhancement; neural loudspeaker control; differentiable mixing; perceptual metrics; and creator interaction. Targeted expansion addressed weak branches found in the earlier audit rather than indiscriminately adding every adjacent generative-audio paper.

The inclusion rule is practical relevance to spatial-audio production or a clearly identified enabling layer. Foundational adjacent work is labeled as such. Generic mono audio generation, all speech enhancement, the full classical acoustics literature, patents, product announcements and every workshop variant are not claimed as exhaustively surveyed. Standards are included for delivery and evaluation context, not counted as academic models. A survey paper is a map and discovery aid, not a substitute for checking the underlying research. [123]

Access depth is recorded beside each paper. A full-text designation indicates that relevant methods, evaluation or limitations were inspected. Where an official abstract, institutional record or author project material was the available source, the judgment is limited to what that material establishes. The contribution paragraphs attribute empirical findings to their authors. The assessment and verdict paragraphs give Spatial9's technical interpretation rather than adding unreported benchmark values.

Version handling is essential in a fast-moving field. Where verified, publication and preprint dates are distinguished. Accepted work is distinguished from a submission or a preprint. Title changes and duplicate arXiv records are not treated as new independent models. The AMBIQUAL conference and journal papers remain separate publications while being identified as one research lineage; the Diff-MST model and its DAW demonstration are separate contributions with different evidence. [117, 118, 5, 6]

The literature cutoff is 25 October 2026. This is a narrative critical review, with an explicit catalog and version history. It makes no claim to preregistered systematic screening, pooled effect sizes or a comprehensive global coverage rate. Related versions and adjacent tasks remain visible so readers can assess the breadth and limits of the synthesis directly.

Spatial9's public descriptions provide context for the company perspective: analysis, placement, scene control, rendering and delivery. Research-preview capabilities retain their stated status. Product descriptions and academic results serve different roles in this paper; the former explain our architectural context, while the latter support the record-specific research findings. Our comparative judgments concern technical fit and implications, with no numerical league table implied. [129, 130, 131]


### About the code examples

Fifteen short excerpts connect selected research findings to inspectable engineering decisions. Each example sits beside its paper review; the links below provide direct access.

Official API examples use public project interfaces checked against the cited sources. Illustrative integration patterns are original explanatory code expressing Spatial9's architectural position, including simplified translations of published equations. They are not extracts from Spatial9's private implementation. Inputs, assumptions and omitted responsibilities appear beside each fragment.


## A task map of AI spatial audio

Spatial audio is a chain of representations and decisions. The same waveform may be analyzed for source identity, transformed into editable objects, rendered for a listener and then compressed for delivery. Research can improve any one of these operations without solving all of them. The table below is a comparison map: it identifies the question each branch answers and the evidence needed to connect it to production.

Inputs matter as much as outputs. A calibrated room mesh, measured listener pose and clean stems supply information that a stereo master does not. We regard both calibrated and minimally informed systems as potentially useful, with different operating contracts. Their results become meaningful when the available information is explicit, including the effort required to obtain it.

Representations also determine what can be edited. Channel-based audio assigns signals to a playback arrangement. Scene-based audio represents a sound field, often through Ambisonics. Object-based audio carries source signals with associated metadata. A renderer can move between representations, but information that has been mixed or discarded is not automatically recovered by changing the file extension. We judge a handoff by the structure actually written to the delivered file. [125, 126]

Physical plausibility and creative appropriateness are separate axes. A film effect can intentionally exceed natural motion; a music mix can favor a stable image over literal source geometry. Spatial9's position is to make the chosen intention explicit and preserve the creator's ability to revise its execution. That allows physical modeling and artistic control to contribute to the same scene without confusing their purposes.

[Image: An object-based mixing session: the place where separation, placement, rendering and delivery decisions meet.]

| Branch | What the research primarily delivers | Question for production |
| Separation and extraction | Sources or stems from mixtures | Are the intended sources complete, stable and suitable for independent control? |
| Localization and understanding | Events, positions, descriptions or answers | Are decisions grounded in acoustic evidence under overlap and changing conditions? |
| Binaural synthesis | Signals for the two ears | Do cues remain convincing across motion, listeners and content? |
| Music and Ambisonic upmixing | Additional channels or a sound-field representation | Does the transformation preserve source identity and support the requested arrangement? |
| Spatial generation | New spatial sound scenes or effects | Are content, position and trajectory controllable without unwanted events? |
| Audio-visual and interactive systems | Sound aligned with visual or dynamic scene context | What happens with off-screen sources, camera motion and user movement? |
| Neural acoustics | Room responses or acoustic fields | How much calibration is required, and what changes invalidate the model? |
| HRTF and binaural enhancement | Listener filters or enhanced ear signals | Does reconstruction improve perceived location, externalization and useful listening? |
| Loudspeaker control | Drive signals or sound-zone filters | Does performance survive real rooms, listener movement and output constraints? |
| Planning and production interaction | Editable parameters, intentions and workflows | Can a user reach and revise the desired result efficiently? |
| Codecs and representations | Compressed signals or learned tokens | Which cues and editing information survive reconstruction? |
| Evaluation and delivery | Metrics, protocols and interoperable structures | Does the final playback preserve the intended behavior? |


## Planning and creative control in production

We judge planning at the point where intention becomes a decision a creator can use. Text, a reference recording and a detected musical event supply different kinds of context. The papers in this chapter show several routes from those inputs to parameters or actions. Our shared criterion is whether the result remains understandable and revisable when the creator changes the brief.

DDSP and differentiable mixing establish a lineage of learned control through structured processing; STASE connects semantic planning with spatial synthesis; S3A places current work within a longer history of object-audio systems. We regard explicit parameters as important prior art. Our position focuses on their meaning within a complete spatial session: the identity of the source, the interval affected, the allowed movement and the delivered representation. [2, 3, 7, 1]

For Spatial9, the first result and the next revision belong to one production experience. A useful system supports acceptance, local correction and recovery while keeping the rest of the arrangement coherent. The workflow studies reinforce the diversity of creator needs. We interpret them as design evidence for progressive control, without converting their exploratory samples into universal adoption claims. [9, 10]

[Image: The production chain this review uses to place every paper. Each stage is judged by whether its information stays useful to the next one.]


### R16: S3A object pipeline

An Audio-Visual System for Object-Based Audio: From Recording to Listening (2018 | IEEE Transactions on Multimedia 20(8), 1919–1931). Status: Journal publication; metadata and associated dataset verified. Reviewed: https://cvssp.org/data/s3a/public/AudioVisualSystem/. Access: Institutional metadata and associated experiment dataset; detailed manuscript review not completed.

Research contribution. The S3A paper documents an audiovisual object-audio system spanning recording through listening. Its institutional record and associated experimental dataset establish relevant prior work on complete object-based workflows, including capture and reproduction. Detailed full-text retrieval was blocked in this review. Consequently, this entry credits the documented system contribution without making algorithm-specific claims about uninspected implementation details. [1]

Spatial9 assessment. We place Spatial9 within the established history of object-audio systems. Persistent objects, adaptable reproduction and integrated workflows are important prior ambitions, and they deserve explicit credit. Our company perspective emphasizes the operational connection between the material a creator already has, the decisions they make and the experience they deliver. That perspective makes the complete workflow the relevant unit of product value.

Our verdict. S3A is essential context for claims about integrated object-based production. It supports a lineage argument, not an artificial numerical contest between unlike systems. For Spatial9, object identity has value when it remains meaningful through revision and handoff. We distinguish an internally editable object from the representation ultimately delivered, because the information available at those stages determines what the next user can change.


### NROOT01: DDSP

DDSP: Differentiable Digital Signal Processing (2020 | ICLR 2020). Status: Conference paper; official research page links proceedings. Reviewed: arXiv 2001.04643v1. Access: Full text: architecture, synthesis components and experimental scope.

Research contribution. DDSP integrates differentiable signal-processing components with neural control. The paper's synthesis experiments use monophonic sources and interpretable oscillators, filters and reverberation. Its relevance here is architectural: learned decisions can operate through structured signal-processing modules. It is foundational adjacent work, not a complete spatial music production system or a demonstration that arbitrary audio transformations preserve an original recording. [2]

Spatial9 assessment. We regard DDSP as foundational prior art for learned control over structured audio processing. Its architectural significance is greater than any claim that adding DSP automatically creates transparency. Meaningful controls still require stable mappings and a clear relationship to the user's intention. We also distinguish synthesis from retaining an existing source; structured processing can support either, with different obligations.

Our verdict. DDSP provides a compelling architectural foundation for interpretable audio systems, while leaving spatial production and source-preservation questions to their respective implementations. Spatial9's position is to make the decision layer and signal path understandable at the level a creator can act on. We value explicit structure when it helps localize a change, preserve an exception or recover a previous decision. The presence of DSP alone is not the product claim.


### Code example 1: Keep a learned control connected to an explicit DSP operation

Illustrative integration pattern. Related reviews: NROOT02, NROOT05. DDSP demonstrates the value of differentiable, interpretable signal operations. Here, a learnable scalar controls an equal-power stereo pan. The signal path remains understandable, and an audio loss can send a gradient back to the parameter that a creator ultimately edits. [2, 3, 136]

```python
import torch
def render_pan(mono, raw_pan):
    angle = 0.5 * torch.pi * torch.sigmoid(raw_pan)
    gains = torch.stack((torch.cos(angle), torch.sin(angle)))
    return mono[:, None] * gains[None, :]
# mono: [samples]; target: [samples, 2], same device/dtype.
raw_pan = mono.new_tensor(0.0, requires_grad=True)
stereo = render_pan(mono, raw_pan)
loss = (stereo - target).square().mean()
loss.backward()
```

Reading the example. The sigmoid bounds the pan between the two speakers. The squared gains sum to one, while autograd exposes the change in loss with respect to raw_pan. This is the architectural point we carry from differentiable DSP: a learned decision can remain an inspectable parameter.

Scope. This original PyTorch example is not the DDSP library's synthesis model. It is a static stereo primitive, not a 3D renderer, and waveform error is only an illustrative objective. Perceptual spatial losses and time-varying control smoothing are separate choices.


### NROOT02: Differentiable mixing console

Automatic Multitrack Mixing with a Differentiable Mixing Console of Neural Audio Effects (2021 | IEEE ICASSP 2021; preprint first posted 2020). Status: Conference publication. Reviewed: arXiv 2010.10291v1. Access: Full text: model, data splits, listening evaluation and discussion.

Research contribution. This work predicts readable mixing controls through a differentiable console with neural effect models. It studies drum recordings and multitrack music, including a listening evaluation with experienced engineers. The music results vary across passages, and training configurations restrict track counts for memory reasons even though the architecture is not intrinsically fixed to one count. It provides early evidence for learned control with subsequent human adjustment. [3]

Spatial9 assessment. We regard editable parameter prediction as an established and valuable research direction. This paper is especially relevant because it connects learned mixing with controls an engineer can understand. Our spatial perspective adds the need to describe object identity, time-dependent position and rendering relationships. Those requirements extend the production context without diminishing the contribution of an adjustable mixing console.

Our verdict. The differentiable console is a strong precedent for human refinement after automated decisions. For Spatial9, the decisive product issue is whether the exposed state corresponds to the change a creator wants to make. Initial quality and revision quality belong to the same workflow. We treat an editable parameter interface as meaningful architecture, while reserving judgments about scale and musical usefulness for the conditions the research actually establishes.


### NROOT03: DeepAFx-ST

Style Transfer of Audio Effects with Differentiable Signal Processing (2022 | Journal of the Audio Engineering Society 70(9), 708–721). Status: Journal publication; DOI 10.17743/jaes.2022.0025. Reviewed: Published 2022 article and arXiv 2207.08759v1. Access: Full text: differentiation methods, experiments and user-interaction discussion.

Research contribution. DeepAFx-ST learns effect parameters from an input recording and a style reference, using differentiable processing and self-supervised training. The paper compares differentiation strategies and studies speech and music. Its discussion distinguishes computational and implementation tradeoffs, including approaches that use the original DSP at inference. It also describes user interaction through reference examples and editable effect parameters. [4]

Spatial9 assessment. We see reference audio as a powerful but incomplete expression of creative intent. A producer can want the reference's tonal balance while rejecting its width, motion or room character. DeepAFx-ST's explicit effect parameters make this ambiguity more tractable because the desired intervention has an identifiable target. Our position is to separate the dimensions of a reference that are requested from those the creator wants to retain.

Our verdict. DeepAFx-ST is a useful reference for controllable style-conditioned processing. Its strongest production implication is the connection between an example, a parameter decision and an editable result. For Spatial9, we favor selective interpretation of references over treating a reference as an indivisible instruction. Spatial arrangement and effect style need a shared brief, with boundaries that keep a local request from becoming a wholesale reinterpretation of the work.


### NROOT05: Diff-MST

Diff-MST: Differentiable Mixing Style Transfer (2024 | ISMIR 2024). Status: Conference publication; author repository and demonstration manuscript confirm venue. Reviewed: arXiv 2407.08889v1. Access: Full text: framework, training and comparative evaluation.

Research contribution. Diff-MST combines a differentiable mixing console, transformer controller and production-style loss to estimate editable parameters from tracks and a reference mix. Its evaluation includes selected music cases, and the authors acknowledge a disadvantage when adapting a mix-to-mix baseline to raw-track inputs. This is an especially relevant example of why benchmark input conditions must be inspected before interpreting a reported comparison. [5]

Spatial9 assessment. We regard Diff-MST as significant prior art for editable style-conditioned mixing. Its baseline discussion also reinforces an essential comparison rule: an adaptation can change what a reported result means. A shared production task and an entire product workflow are both valuable units of assessment, but they answer different questions. Spatial9's view keeps the input contract visible before drawing a comparative conclusion.

Our verdict. Diff-MST is a strong reference for controllable mixing assistance and an instructive example of benchmark interpretation. Its relevance to Spatial9 lies in explicit decisions, reference sensitivity and continuity across a session. Additional spatial functions belong in their own assessment rather than being used to discount a mix-style model. We place value on the creator's ability to refine the resulting controls, with the scope of published evidence kept clear.


### NROOT06: Diff-MSTC

Diff-MSTC: A Mixing Style Transfer Prototype for Cubase (2024 | ISMIR 2024 demonstration). Status: Demonstration manuscript; related CHI 2025 demo identified but not separately counted. Reviewed: arXiv 2411.06576v1. Access: Full text: prototype workflow and stated future evaluation.

Research contribution. Diff-MSTC integrates mixing-style transfer into Cubase and lets users refine predicted console parameters. The reviewed prototype accepts up to twenty tracks and asks the user to select input and reference sections. The manuscript explicitly defers formal user-experience studies. It demonstrates integration and interaction, while leaving broader usability effectiveness as a research question. [6]

Spatial9 assessment. We consider workflow integration a substantive technical contribution. A model's output enters an existing session with manual decisions, reference choices and familiar tools. Diff-MSTC matters because its interaction occurs inside that environment. Our position treats handoff, correction and retained session state as part of the system boundary. A polished isolated render addresses only one portion of that production relationship.

Our verdict. Diff-MSTC is an important integration precedent whose prototype evidence remains distinct from established usability outcomes. For Spatial9, the lesson is to judge assistance by its fit with the creator's next action. Save and reload, local revision and preservation of deliberate choices are core concerns in that judgment. We value the paper for making the interaction concrete, while keeping its deferred user-experience questions open.


### R01: STASE

STASE: A spatialized text-to-audio synthesis engine for music generation (2025 | LLM4Music workshop at ISMIR 2025). Status: Workshop acceptance stated in author manuscript. Reviewed: arXiv 2509.11124v1. Access: Full text: methods, implementation, results, discussion.

Research contribution. STASE organizes text or supplied mono stems into a spatial plan, then renders that plan using signal processing. Its workshop manuscript describes language-model planning, retrieval and spatial effects, while its results are demonstrations rather than a controlled comparative benchmark. The authors acknowledge difficulties with dense or reverberant source mixtures. The relevant contribution is an explicit connection between semantic decisions and an executable audio pipeline. [7]

Spatial9 assessment. We regard STASE as direct architectural prior art for semantic planning followed by spatial rendering. Its most useful idea is the explicit plan between an instruction and a signal-processing action. That separation gives a production system a place to represent intent, expose a decision and support correction. Our reservation concerns the completeness of the authoring contract: an executable plan alone does not specify how source identity, exceptions and revisions remain coherent throughout a session.

Our verdict. STASE is a relevant planning reference with a clear place in the history of hybrid spatial systems. For Spatial9, the important architectural commitment is a persistent, inspectable scene whose decisions remain meaningful after the first render. We value language-driven planning when it produces precise scene operations. We judge the resulting workflow by the creator's ability to change one intention while retaining the rest of the arrangement, rather than by the fluency of the generated plan.


### Code example 2: Validate a spatial instruction before changing the scene

Illustrative integration pattern. Related reviews: R16. STASE makes a structured spatial plan an explicit stage of generation. Our position is that an instruction becomes useful when its scope is executable and reversible. This fragment applies a position-only edit to an existing object while retaining its source identity and every unrelated scene field. [7, 1]

```python
from copy import deepcopy
from math import isfinite
def apply_position(scene, object_id, position):
    fields = ("azimuth_deg", "elevation_deg", "distance_m")
    if set(position) != set(fields):
        raise ValueError("Position fields do not match the schema")
    az, el, radius = (float(position[k]) for k in fields)
    if not all(map(isfinite, (az, el, radius))):
        raise ValueError("Non-finite position")
    if not (-180 <= az <= 180 and -90 <= el <= 90):
        raise ValueError("Direction outside the declared range")
    if radius <= 0:
        raise ValueError("Distance must be positive")
    updated = deepcopy(scene)
    updated["objects"][object_id]["position"] = dict(
        zip(fields, (az, el, radius)))
    return updated
```

Reading the example. The illustrative scene convention uses degrees, metres, forward azimuth zero, positive azimuth toward the left and positive elevation upward. A missing object raises an error instead of creating a new source. The returned copy permits comparison or undo without rerunning synthesis.

Scope. This is an application-side contract, not STASE's published API. It omits schema parsing, concurrent-edit handling and renderer-specific distance behaviour; valid coordinates alone do not establish perceived location.


### R02: SEND

SEND: A Spatial Event Neural Detector for Intentional Object Motion in Immersive Music Mixing (2026 | DAFx 2026). Status: Conference proceedings. Reviewed: https://dafx26.mit.edu/assets/papers/DAFx26_paper_09.pdf. Access: Full text: architecture, dataset, evaluation, conclusion.

Research contribution. SEND learns when an object should move from a target track and its musical background. The DAFx paper uses professional immersive projects to study intentional event timing. It does not claim to synthesize a complete continuous three-dimensional trajectory; that extension is future work. The authors also discuss occasional isolated activations and spatial jitter. This places the work in musical motion decision making, upstream of trajectory design and rendering. [8]

Spatial9 assessment. We consider SEND important because it assigns technical substance to musical restraint. Movement has an appropriate time and context; constant activity is not a useful definition of immersion. Our position is that an intentional event, its musical purpose and its permitted movement need separate representations. This makes a stationary choice as legitimate as an active one and gives the creator a meaningful way to override automation.

Our verdict. SEND is a strong conceptual reference for motion decisions, with a deliberately narrower scope than complete trajectory authoring. Spatial9's spatial-control philosophy treats event timing and path design as separate responsibilities. A well-timed motion opportunity still needs a suitable path, boundaries and continuity. We credit SEND for clarifying the first responsibility and regard the connection to an editable musical scene as the central production implication.


### Code example 3: Turn a movement window into a continuous trajectory

Illustrative integration pattern. SEND identifies when movement is appropriate. The renderer still needs a trajectory with defined behaviour before, during and after that interval. This fragment converts one accepted movement window into a smooth Cartesian transition between two positions in metres. [8]

```python
import numpy as np
def transition(t_s, start_s, stop_s, xyz_a, xyz_b):
    if stop_s <= start_s:
        raise ValueError("Movement needs a positive duration")
    t = np.asarray(t_s, dtype=float)
    a, b = np.asarray(xyz_a, float), np.asarray(xyz_b, float)
    if a.shape != (3,) or b.shape != (3,):
        raise ValueError("Positions must be 3D vectors")
    u = np.clip((t - start_s) / (stop_s - start_s), 0.0, 1.0)
    weight = u * u * (3.0 - 2.0 * u)
    return a + weight[..., None] * (b - a)
```

Reading the example. The object remains at its first position before the window and its destination afterwards. The cubic easing has zero endpoint velocity. This expresses our separation of a musical movement decision from the curve that executes it.

Scope. This original interpolation is not SEND's detector or a claimed reproduction of its postprocessing. Inputs are finite and use one coordinate frame. Combining windows, velocity limits and paths around obstacles require additional controls.


### NROOT04: AI mixing adoption study

Adoption of AI Technology in the Music Mixing Workflow: An Investigation (2023 | 154th Audio Engineering Society Convention). Status: Convention research paper; author manuscript reviewed. Reviewed: arXiv 2304.03407v2, 8 September 2023. Access: Full text: interviews, questionnaire, forum analysis and findings.

Research contribution. This mixed-method study combines five expert interviews, a questionnaire involving twenty-two experienced engineers and forum analysis. It distinguishes amateur, pro-am and professional expectations of AI mixing tools, including different preferences for automation and control. The sample and recruitment methods support exploratory design insights; they do not establish market-wide adoption rates or a universal preference shared by every creator. [9]

Spatial9 assessment. We take the variation in user expectations seriously. A useful starting point, a precise intervention and detailed manual control are different needs that can arise even within one person's session. Our position is progressive control: a coherent first result accompanied by access to the decisions that matter for revision. The study provides design insight into that position, without supplying market-wide adoption estimates.

Our verdict. This research is more valuable as a guide to product behavior than as a source of market claims. Spatial9's perspective accommodates both rapid assistance and deliberate intervention. We consider acceptance, correction and rejection legitimate user actions, and judge automation by how well it supports them. A creator retains agency when the system is predictable and the relevant decisions remain recoverable, without requiring constant inspection of every internal operation.


### NROOT07: Professional workflow ethnography

It's All About Speed: AI's Impact on Workflow in Music Production (2025 | AES International Conference on Machine Learning and Artificial Intelligence for Audio 2025; preprint posted 2026). Status: Conference presentation stated in primary manuscript metadata. Reviewed: arXiv 2605.29931v1, 28 May 2026. Access: Full text: ethnographic method, participant scope and discussion.

Research contribution. This ethnographic study examines five professional producers and engineers through studio observation or interviews. It discusses speed, trust, controllability and creative agency, including differing reactions to opaque tools. Its local, small qualitative sample provides detailed workflow observations rather than a population estimate. The conference year and later preprint date are kept separate in this review. [10]

Spatial9 assessment. We interpret speed as progress toward an accepted result, including the cost of correction and recovery. The most transparent interface is not automatically the best one for every moment in a session. Some actions benefit from immediate assistance; others demand direct intervention. Our position is to make control available where it affects the decision, with predictable outcomes and recoverable edits.

Our verdict. This study provides useful qualitative evidence for the relationship between speed, trust and agency. It does not establish a universal creator preference, and our product interpretation respects that variation. For Spatial9, time to the first render and time to an acceptable revision are different aspects of usefulness. We value automation that reduces the total burden on the creator, including the work required when an initial suggestion is rejected.


## Source separation and stable production assets

We regard source recovery as the beginning of an immersive production contract. The separator determines which material becomes available, while the authoring system determines how that material is identified, positioned, and revised. Broad stems, prompted targets, and discovered sources are useful outputs with different granularity and completeness. Our position is that these differences belong in the asset model. A creator needs to know whether an apparent object is one instrument, a grouped stem, or a candidate extraction whose identity still needs resolution.

We favor separation architecture that preserves provenance and makes model substitution understandable. Recovery quality, inventory completeness, and persistent identity deserve distinct judgments because success on one does not settle the others. The papers in this chapter support several valuable routes into the recording, from reproducible baselines to open-vocabulary selection. For Spatial9, the product implication is stable control over the resulting assets. A source becomes production-ready through a clear identity and revision contract, not merely because a model can output its waveform.

[Image: Separation as an asset boundary: named stems, a prompted target with its residual, and a scene identity that survives waveform revisions (code examples 4, 5 and 6).]


### SEP01: Hybrid Transformer Demucs

Hybrid Transformers for Music Source Separation (2023 | ICASSP 2023). Status: Peer-reviewed conference paper. Reviewed: arXiv:2211.08553v1. Access: Full text: architecture, training, results, inference procedure.

Research contribution. Hybrid Transformer Demucs connects temporal and spectral representations through cross-domain attention for music source separation. Its principal production output is a set of broad stems, commonly vocals, drums, bass, and other. The paper evaluates separation quality and examines the contribution of architecture, training data, and inference procedures. It provides an important upstream component for workflows that need access to material embedded in a stereo recording. [11]

Spatial9 assessment. We regard HTDemucs as an important recovery layer for immersive production. Its broad stem categories are useful starting assets, but they do not define every object a producer may want to position. The 'other' category makes that distinction especially clear. We also attach reported performance to its data and inference configuration: architectural reputation alone is insufficient grounds for choosing a checkpoint. The relevant production value is usable access to the recording's constituent material. [11]

Our verdict. HTDemucs belongs near the beginning of a spatial workflow, with separation and scene design treated as distinct responsibilities. A recovered stem can support several creative placements, while a coarse stem may need further decomposition before independent placement makes sense. We favor architectures that retain the original recording, identify derived assets, and expose the remaining granularity limit to the creator. That gives a strong separator a clear role without making its output taxonomy dictate the entire immersive mix.


### Code example 4: HTDemucs and the named stem output contract

Official API example. Related reviews: SEP10. We treat separation as a named asset boundary. This compact example uses the official repository API; it assumes that checkout and its dependencies are installed, weights are accessible, and mix.wav exists. [11, 137, 138]

```python
from pathlib import Path
from demucs.api import Separator
separator = Separator(model="htdemucs", device="cpu", shifts=0)
mixture, stems = separator.separate_audio_file(Path("mix.wav"))
sample_rate = separator.samplerate
vocals = stems["vocals"]  # [channels, samples] at model sample rate
```

Reading the example. We retain the returned mixture and sample rate alongside named stems. Downstream code selects vocals by name, rather than assuming a tensor index or reusing the input file rate. shifts=0 disables random-shift averaging for this compact path.

Scope. The repository API is the reference, not a promise about every PyPI release. htdemucs is the named checkpoint, not the fine-tuned ensemble or every configuration reported in the paper. The fragment omits export, loudness policy, object identification and rendering; no performance result is implied.


### SEP02: Open-Unmix

Open-Unmix - A Reference Implementation for Music Source Separation (2019 | Journal of Open Source Software 4(41), 1667 (2019)). Status: Peer-reviewed software paper. Reviewed: Published JOSS paper. Access: Full text.

Research contribution. Open-Unmix establishes a reproducible reference implementation for music source separation using spectrogram estimation, a bidirectional recurrent model, and waveform reconstruction. The software paper emphasizes accessible components and a clear experimental baseline. Its importance extends beyond a particular score: it makes a complete separation procedure available for inspection, substitution, and comparison by other researchers. [12]

Spatial9 assessment. We value Open-Unmix for making the separation procedure understandable and reproducible. A reference implementation has lasting engineering value even when a newer model posts a higher score. Its usefulness is strongest where the team needs to trace an output back to a known checkpoint, preprocessing path, and reconstruction procedure. We do not equate that clarity with a complete production environment; we regard it as a sound foundation for assembling one responsibly. [12]

Our verdict. Reproducibility is a product requirement, and Open-Unmix is a useful reference for that requirement. We favor a separation interface that permits model substitution while preserving the surrounding asset and rendering contracts. A platform becomes easier to maintain when a changed separator has an identifiable effect rather than silently changing every later stage. For Spatial9, the enduring lesson is that traceable processing and understandable dependencies deserve a place alongside perceptual quality in architecture decisions.


### SEP03: Spleeter

Spleeter: a fast and efficient music source separation tool with pre-trained models (2020 | Journal of Open Source Software 5(50), 2154 (2020)). Status: Peer-reviewed software paper. Reviewed: Published JOSS paper. Access: Full text.

Research contribution. Spleeter packages pretrained source-separation models behind an accessible interface, including configurations producing two, four, or five stems. Its paper describes high processing throughput and evaluates separation using several distortion measures. It helped make source extraction practically usable without requiring users to train a model or assemble a research pipeline. [13]

Spatial9 assessment. We regard Spleeter's accessible pretrained tooling as an important contribution to practical audio work. We also take its buffering distinction seriously: fast batch throughput and responsive live interaction are different promises. The paper includes spatial distortion measures, so reducing its evaluation to SDR alone gives an inaccurate account. Our production assessment is therefore specific: Spleeter is relevant to file-based source preparation, while a live workflow has additional timing obligations beyond its headline processing speed. [13]

Our verdict. Spleeter's strongest role is a straightforward extraction stage in an offline workflow. The architecture implication is to keep turnaround time, first playable output, and sustained interaction delay as separate product commitments. A creator preparing a catalog can value batch speed differently from a performer monitoring a live signal. We favor interfaces that make that distinction visible and communicate the chosen stem configuration clearly, rather than allowing one speed figure to define suitability for every use case.


### SEP04: Band-split RNN / BSRNN

Music Source Separation with Band-split RNN (2023 | IEEE/ACM Transactions on Audio, Speech, and Language Processing 31, 1893–1901 (2023)). Status: Peer-reviewed journal paper. Reviewed: arXiv abstract and IEEE bibliographic record. Access: Primary abstract and manuscript introduction inspected; detailed limitations not independently reviewed.

Research contribution. Band-split RNN organizes a music spectrogram into subbands and models dependencies over both frequency bands and time. The architecture reflects the differing spectral structure of musical sources rather than applying one undifferentiated sequence model to the entire spectrum. Its reported contribution is improved music separation within the evaluated data and training configuration. [14]

Spatial9 assessment. We find BSRNN's use of spectral structure persuasive as an architectural idea: musical sources need not be modeled as if every frequency region has the same role. That is a reason to take the design seriously, not a basis for assuming superiority on every instrument. Our assessment rests on the verified architecture and publication scope. The inspected material does not support a more detailed claim about unusual instruments, deployment cost, or particular audible failure modes. [14]

Our verdict. BSRNN is a credible separation component whose value depends on the material and operating configuration. We favor selecting that component by the quality of the recovered assets required by a production, rather than by architectural fashion. Once stems are positioned independently, the authoring system also needs a clear account of what each stem contains. For Spatial9, band-aware modeling is useful upstream engineering; source identity, placement, and revision remain distinct downstream responsibilities.


### SEP05: BS-RoFormer

Music Source Separation with Band-Split RoPE Transformer (2024 | ICASSP 2024). Status: Peer-reviewed conference paper. Reviewed: arXiv:2309.02612v2. Access: Full text methods and evaluation inspected, including training-data regimes and overlap reconstruction.

Research contribution. BS-RoFormer applies transformer processing with rotary positional embeddings to a band-split music representation. The paper evaluates a challenge system using additional training songs and smaller configurations on MUSDB18HQ. Its implementation predicts complex masks and models dependencies across time and bands, providing a strong separator family for controlled downstream comparisons. [15]

Spatial9 assessment. We regard BS-RoFormer as a strong separator family, with the important qualification that its training-data regimes are part of the result. A challenge system with additional songs and a smaller MUSDB configuration cannot be treated as interchangeable evidence. The production question is also broader than aggregate separation quality: residual material acquires a spatial consequence when a stem is moved. That consequence makes the selected checkpoint and downstream use inseparable from a meaningful recommendation. [15]

Our verdict. BS-RoFormer deserves a serious place in source-recovery architecture, while its role remains recovery of usable stems. We favor retaining the checkpoint, processing settings, and asset lineage with those stems so later revisions have a stable reference. A separation score can justify confidence in the recovery stage; it cannot determine the intended layout or the acceptability of a particular immersive presentation. For Spatial9, the sensible design is a strong separator behind an explicit, editable scene contract.


### SEP06: Mel-Band RoFormer

Mel-Band RoFormer for Music Source Separation (2023 | arXiv preprint; described as submitted to ISMIR 2023 late-breaking/demo track). Status: Main-conference peer review not verified. Reviewed: arXiv:2310.01809v1. Access: Full text architectural comparison and source-wise experimental results inspected.

Research contribution. Mel-Band RoFormer replaces the non-overlapping band partition in BS-RoFormer with overlapping mel-frequency bands. The reviewed paper tests smaller models on MUSDB18HQ without extra training data. It reports improvements for several source categories, while the bass result is less successful than the preceding band-split arrangement. This makes the paper useful evidence that representation choices can affect sources differently. [16]

Spatial9 assessment. We find the source-dependent outcome more useful than a blanket claim that mel bands are better. The reviewed configuration improves several categories while treating bass less successfully, so the representation change has a tradeoff worth retaining in the discussion. We attach that judgment to this paper's configuration. A later community checkpoint with different training material is a different selection decision, even when it carries the same model-family name. [16]

Our verdict. Mel-Band RoFormer argues for source-aware model selection rather than a universal architecture winner. We favor a production decision that respects the weakest important source in the actual arrangement, especially when that source will become spatially exposed. A vocal-led demonstration cannot settle the suitability of a bass-driven mix. For Spatial9, the practical implication is to treat the checkpoint and the recovered asset as the unit of choice, with the model family providing context rather than a quality guarantee.


### SEP07: SAM Audio

SAM Audio: Segment Anything in Audio (2026 | ICML 2026, Proceedings of Machine Learning Research 306, 112167–112194). Status: Peer-reviewed conference paper. Reviewed: arXiv:2512.18099v1 full text plus ICML 2026 publication record. Access: Full text: prompting, architecture, evaluation, latency, limitations.

Research contribution. SAM Audio combines text, visual, and temporal prompting in a general source-separation framework. It returns a selected target and residual audio and includes a Multi-diffusion procedure for long-form inference. Its ICML 2026 publication extends the interaction beyond fixed instrument classes and offers a relevant component for selecting material in complex audiovisual scenes. [17]

Spatial9 assessment. We see SAM Audio's major value in flexible access to a desired sound through several forms of prompting. Its long-form inference deserves explicit credit. The remaining production distinction is between selecting a target and maintaining an inventory: repeated extractions do not, by themselves, define which assets are unique, complete, or persistent. We regard that as a system-design responsibility around the model, rather than evidence that a prompted extraction method has failed at its intended task. [17]

Our verdict. SAM Audio is a compelling semantic access layer for complex material, with an asset-management layer needed around each accepted extraction. We favor stable source identifiers, retained residuals, and explicit decisions about overlapping selections so a later instruction refers to the same material. Prompt flexibility is valuable precisely because it broadens what a creator can request. For Spatial9, that flexibility earns its place when the resulting selection becomes a dependable editing asset rather than another disconnected output file.


### Code example 5: SAM Audio and a prompted extraction with its residual

Official API example. Related reviews: SEP08, SEP10. We expose the text query while keeping both returned signals available. This uses the official SAM Audio interface and assumes its dependencies, CUDA, authorized checkpoint access and an existing mix.wav. [17, 139, 140]

```python
import torch
from sam_audio import SAMAudio, SAMAudioProcessor
checkpoint = "facebook/sam-audio-large"
model = SAMAudio.from_pretrained(checkpoint).eval().cuda()
processor = SAMAudioProcessor.from_pretrained(checkpoint)
batch = processor(
    audios=["mix.wav"], descriptions=["lead vocals"]
).to("cuda")
with torch.inference_mode():
    result = model.separate(
        batch, predict_spans=False, reranking_candidates=1
    )
target = result.target[0]  # first item in the returned batch list
residual = result.residual[0]
sample_rate = processor.audio_sampling_rate
```

Reading the example. The current implementation returns lists of per-item waveforms, so we select item zero explicitly. We retain the residual as an editing asset instead of silently discarding it. A prompt expresses an extraction request; our application still needs a separate source identity.

Scope. This is one text-conditioned path, with span prediction disabled and one candidate. It omits visual prompts, temporal anchors, orchestration and export. Keeping target and residual does not establish exact reconstruction of the original mixture. This excerpt reports no execution or quality measurement.


### SEP08: AudioSep

Separate Anything You Describe (2025 | IEEE Transactions on Audio, Speech and Language Processing 33, 458–471 (2025)). Status: Peer-reviewed journal paper. Reviewed: arXiv:2308.05037v3, revised December 2024; 2025 journal metadata. Access: Full text: QueryNet/SeparationNet, data conversion, evaluation mixtures.

Research contribution. AudioSep pairs a language-audio representation with a separation network to recover sounds described in text. It broadens target selection beyond a fixed output taxonomy. The reviewed training description converts AudioSet material to mono, and the publication evaluates open-vocabulary separation across several constructed and real-audio settings. The journal publication followed the earlier preprint and appeared in 2025. [18]

Spatial9 assessment. We value AudioSep's open vocabulary because a production rarely fits perfectly into a fixed set of stem names. Its reviewed mono training path also makes the intended fidelity contract important. Recovering a sound for a newly authored scene and retaining the recording's original stereo relationships are different briefs. We regard both as legitimate uses, but the language prompt alone cannot decide which spatial information the creator expects to preserve. [18]

Our verdict. AudioSep is especially relevant when the desired material falls outside conventional stem categories. We favor an extraction workflow that declares whether original spatial cues are part of the asset contract or whether placement will be authored anew. That decision belongs beside the source selection, not hidden inside downstream processing. For Spatial9, semantic access is a strong capability, while source choice, cue preservation, and new spatial placement remain separate decisions with different creative consequences.


### SEP09: SAM Audio Judge

SAM Audio Judge: A Unified Multimodal Framework for Perceptual Evaluation of Audio Separation (2026 | arXiv:2601.19702, 27 January 2026). Status: Preprint; peer-reviewed venue not verified. Reviewed: arXiv:2601.19702v1. Access: Primary full text introduction and evaluation scope inspected; no detailed author limitation claimed.

Research contribution. SAM Audio Judge estimates perceptual separation quality from multimodal context without requiring a clean reference waveform. The paper develops judgments covering extraction precision, recall, faithfulness, and overall quality, and compares them with human ratings. It addresses the practical difficulty of evaluating mixtures for which isolated ground truth is unavailable. [19]

Spatial9 assessment. We regard SAM Audio Judge as useful infrastructure for assessing extraction when a clean reference is unavailable. We do not treat its separation-quality contract as a certificate for an immersive presentation. Source faithfulness and the correctness of a requested trajectory answer different questions. That distinction does not imply that the judge is insensitive to every spatial error; it means its reported remit is insufficient to stand in for all spatial and delivery requirements. [19]

Our verdict. This judge belongs in a quality system as a focused diagnostic, with its output labeled by the property it estimates. We favor keeping extraction quality, spatial intent, and delivery correctness visible as separate judgments rather than compressing them into one unexplained score. For Spatial9, a useful automated assessment tells a creator what needs attention and why. A favorable learned rating is informative about its stated task; acceptance of the complete immersive mix requires a broader product contract.


### SEP10: MuS3D

Music Source Separation via Stem Discovery (2026 | arXiv:2609.30912, 25 September 2026). Status: Preprint; peer-reviewed venue not verified. Reviewed: arXiv:2609.30912v1. Access: Full text: recursive discovery, evaluation, limitations.

Research contribution. MuS3D investigates discovering and separating musical sources without requiring a fixed stem list or a manually supplied description for each target. A recursive embedding-generation stage identifies candidate sources that guide separation. The paper therefore addresses a practical gap between conventional four-stem models and the more varied source inventories needed by many productions. [20]

Spatial9 assessment. We regard automatic stem discovery as closer to a producer's inventory problem than a permanently fixed output list. MuS3D therefore addresses an important missing step. Its reported missed or misclassified sources and unresolved whole-song consistency place a clear boundary around that promise. A collection of individually convincing stems is still incomplete when a musically important source is absent. We give discovery its own value while keeping inventory completeness distinct from the quality of successful extractions. [20]

Our verdict. MuS3D is an important direction for constructing candidate source inventories, with candidate identity requiring explicit acceptance before it becomes a durable editing target. We favor an authoring model that makes uncertain, duplicate, and missing sources visible instead of treating discovery as a finished scene. For Spatial9, the architectural implication is persistent asset identity across the production: a recurring instrument and its saved edits must remain associated even when the discovery process operates on shorter excerpts.


### Code example 6: Keep scene identity across waveform revisions

Illustrative integration pattern. Related reviews: SEP01, SEP07, SEP13. Our integration principle is to separate a logical source from its waveform revision. This original editorial sketch assumes the creator has confirmed that a new extraction represents the same source; all identifiers and paths are illustrative. [20, 24]

```python
from dataclasses import dataclass, replace
@dataclass(frozen=True)
class StemAsset:
    object_id: str
    revision: int
    audio_uri: str
    extraction_run: str
asset = StemAsset("source-7", 1, "stems/v1/vocals.wav", "run-12")
placement = {asset.object_id: {"azimuth_deg": -25.0}}
# Creator-confirmed replacement: same source, different waveform.
replacement = replace(
    asset, revision=2, audio_uri="stems/v2/vocals.wav",
    extraction_run="run-13",
)
assets = {replacement.object_id: replacement}
source_placement = placement[replacement.object_id]
```

Reading the example. The persistent object identifier owns placement; the asset revision owns the waveform; the extraction run records its processing provenance. An accepted replacement therefore keeps its scene placement without treating an output slot, prompt or filename as identity. This is application architecture motivated by the separation literature, not code attributed to MuS3D.

Scope. This sketch does not infer correspondence between sources, resolve duplicates or implement a database. The extraction run needs an external manifest for checkpoint, input and parameters. No private Spatial9 implementation or automatic tracking capability is claimed.


### SEP26: Waveformer

Real-Time Target Sound Extraction (2023 | ICASSP 2023). Status: Peer-reviewed conference paper. Reviewed: arXiv:2211.02250, camera-ready April 2023. Access: Primary abstract and manuscript overview inspected; runtime behavior not reproduced.

Research contribution. Waveformer addresses real-time target sound extraction with a causal architecture combining dilated-convolution encoding and transformer decoding. It targets streaming selection of a desired sound, making it a useful reference when evaluating systems that must operate before a complete recording is available. [21]

Spatial9 assessment. We regard Waveformer as relevant to a genuinely streaming extraction brief because its design is causal. That operating constraint changes the comparison: an offline separator can use future context that a live system does not have. We therefore reject an unrestricted offline score as a sufficient ranking of the two roles. We also distinguish model causality from the delay of the full listening path, which includes other processing and transport responsibilities. [21]

Our verdict. Waveformer is an appropriate architectural reference for low-delay target extraction. We favor defining the live signal path as a whole, with a clear account of buffering, execution, rendering, and playback. For Spatial9, its relevance is the discipline imposed by a causal operating contract: every dependent stage must respect the same timing promise. An offline preparation workflow can make a different quality-time tradeoff. Those are product choices to state explicitly, rather than collapse into one general claim of model superiority.


## Spatial detection, localization and reference data

We treat detected geometry as evidence about a recording and authored geometry as an expression of intent. That distinction lets the same perception method serve faithful reconstruction and deliberate creative placement. It also makes the meaning of each output essential: activity is not distance, a direction is not a persistent identity, and an estimated physical position is not automatically the producer's preferred presentation. Our architectural position is to retain those meanings explicitly through every handoff into a scene.

We value controlled scenes and real recordings for different reasons. Controlled evidence makes particular relationships interpretable; real-scene data connects the discussion to actual capture conditions. Neither alone defines the acceptance criteria for every musical production. For Spatial9, the useful design is an inspectable relationship between observations, source associations, and creator choices. This preserves research contributions at their stated scope while making clear which additional responsibilities belong to the production interface rather than to the detector or dataset.

[Image: ACCDOA couples event activity to vector length and direction to orientation (code example 7).]


### SEP11: SELDnet

Sound Event Localization and Detection of Overlapping Sources Using Convolutional Recurrent Neural Networks (2019 | IEEE Journal of Selected Topics in Signal Processing 13(1), 2019; early online 2018). Status: Peer-reviewed journal paper. Reviewed: arXiv:1807.00129v3 abstract and publication metadata. Access: Primary abstract and manuscript introduction inspected; detailed limitations not independently reviewed.

Research contribution. SELDnet jointly estimates sound-event activity and three-dimensional direction of arrival from multichannel recordings through a convolutional recurrent architecture. It established a practical formulation for connecting recognition of overlapping events with their acoustic directions. The paper evaluates several array and acoustic conditions, making it foundational to spatial perception rather than a system for composing an immersive music mix. [22]

Spatial9 assessment. We regard SELDnet as foundational perception work because it connects an event with a direction in the recording. Our production assessment starts with the meaning of that estimate: observed location is evidence, while intended placement is a creative decision. Copying a detected direction can be appropriate for reconstruction and inappropriate for a deliberate remix. We therefore value the output as scene information, with its array context retained, rather than as an automatic rule for musical placement. [22]

Our verdict. Localization belongs in the evidence layer of a production system. We favor an explicit distinction between recorded coordinates and authored coordinates, with the creator's choice retained when they differ. That makes the same perception method useful for both faithful reconstruction and intentional rearrangement. For Spatial9, SELDnet's lasting contribution is the joint event-and-direction formulation; the architectural responsibility above it is deciding how acoustic evidence informs a scene without confusing accuracy of observation with artistic authority.


### SEP12: ACCDOA

ACCDOA: Activity-Coupled Cartesian Direction of Arrival Representation for Sound Event Localization and Detection (2021 | ICASSP 2021). Status: Peer-reviewed conference paper. Reviewed: arXiv:2010.15306, revised February 2021. Access: Primary representation definition inspected; full manuscript retrieved.

Research contribution. ACCDOA combines activity and direction in one Cartesian output representation for sound-event localization and detection. The orientation of the vector represents direction of arrival, while its magnitude expresses source activity. This simplifies the relationship between detecting an event and estimating its location within the model's stated task. [23]

Spatial9 assessment. We find ACCDOA elegant because it makes activity and direction part of one representation. Its magnitude, however, represents activity rather than distance. That definition is decisive for integration: a mathematically valid vector can still produce a wrong scene if another component reads it as a point in metres. We regard this as a lesson in interface semantics, not a weakness of the representation. A compact output is useful only when its meaning survives the handoff. [23]

Our verdict. ACCDOA is a useful activity-and-direction interface, with physical distance requiring a separately defined quantity. We favor typed scene attributes that distinguish direction, activity, confidence, and radius instead of treating every three-number output as interchangeable geometry. For Spatial9, this is a concrete architectural rule: model outputs must enter a scene through a documented interpretation of their units and reference frame. The simplicity of the encoder representation is valuable; preserving its meaning makes that simplicity safe to use.


### Code example 7: Decode ACCDOA activity and direction separately

Illustrative integration pattern. Related reviews: SEP13. We decode the representation before turning predictions into scene controls. This original NumPy fragment assumes single-ACCDOA predictions arranged as [frames, classes, xyz] and a caller-supplied activity threshold. [23, 24]

```python
import numpy as np
def decode_accdoa(xyz, threshold):
    xyz = np.asarray(xyz, dtype=np.float64)
    if xyz.ndim != 3 or xyz.shape[-1] != 3:
        raise ValueError("Expected [frames, classes, xyz].")
    if not np.isfinite(threshold) or threshold <= 0:
        raise ValueError("Use a finite, positive activity threshold.")
    if not np.isfinite(xyz).all():
        raise ValueError("Predictions must be finite.")
    activity = np.linalg.norm(xyz, axis=-1)
    active = activity > threshold
    direction = np.full_like(xyz, np.nan)
    np.divide(xyz, activity[..., None], out=direction,
              where=active[..., None])
    return activity, active, direction
```

Reading the example. ACCDOA couples activity to vector length and direction to orientation. We normalize only active vectors; inactive directions remain explicitly undefined. The vector norm is neither source distance nor a calibrated probability. This preserves the distinction between detecting an event and assigning it a geometric direction.

Scope. The threshold is an application choice. Directions remain in the model coordinate frame; conversion to a world or listener frame is separate. This does not implement a network, Multi-ACCDOA duplicate merging, source tracking or distance estimation.


### SEP13: Multi-ACCDOA

Multi-ACCDOA: Localizing and Detecting Overlapping Sounds from the Same Class with Auxiliary Duplicating Permutation Invariant Training (2022 | ICASSP 2022). Status: Peer-reviewed conference paper. Reviewed: arXiv:2110.07124v2. Access: Full text: output representation, ADPIT, merging and experiments.

Research contribution. Multi-ACCDOA extends activity-coupled direction estimates to overlapping sources of the same event class. Auxiliary duplicating permutation-invariant training supports multiple output tracks, while an angular merging procedure consolidates similar predictions. The reviewed experiments use a track capacity chosen for the dataset's overlap conditions. [24]

Spatial9 assessment. We regard Multi-ACCDOA's treatment of same-class overlap as a material advance over a representation that cannot express those simultaneous sources. The separate production issue is identity over time. A detector's flexible track assignments and a creator's persistent object names have different jobs. We do not infer an identity failure from permutation-invariant training; we conclude that a saved edit needs an explicit association policy beyond the detector's output slots. [24]

Our verdict. Multi-ACCDOA is useful for representing overlapping detections, with a persistent identity layer needed before those detections become editable objects. We favor keeping track capacity and association decisions visible in the scene model. Two simultaneous guitars can share a class while requiring different names, trajectories, and revisions. For Spatial9, the crucial architecture choice is to bind an edit to the intended source identity, rather than to whichever prediction slot happens to carry that source at a particular instant.


### SEP14: EINV2

An Improved Event-Independent Network for Polyphonic Sound Event Localization and Detection (2021 | ICASSP 2021). Status: Peer-reviewed conference paper. Reviewed: arXiv:2010.13092, revised February 2021. Access: Primary abstract reviewed and full manuscript retrieved; no detailed author limitation asserted.

Research contribution. EINV2 develops an event-independent approach to polyphonic sound-event localization and detection, using attention, parameter sharing, and permutation-invariant training. Its outputs organize overlapping events into tracks with associated detection and directional information. The method addresses coordination between the recognition and localization tasks in a multichannel recording. [25]

Spatial9 assessment. We value EINV2's event-independent organization as a way to coordinate recognition and localization in polyphonic scenes. Our production judgment concerns the meaning of those tracks, not an unverified defect in the detector. A prediction slot is an operational output; an editable asset carries a continuing identity and history. We regard that distinction as essential when a source disappears, returns, or overlaps with another event of the same class. [25]

Our verdict. EINV2 is relevant to the perception stage, while a production scene needs a distinct policy for converting event tracks into persistent assets. We favor retaining the observation and its association separately so a correction to one does not silently rewrite the other. For Spatial9, that separation makes an editing system more understandable: recognition errors, mistaken source associations, and creative placement decisions remain different causes with different remedies. The detector contributes evidence; the scene model carries the enduring editing contract.


### SEP15: STARSS23

STARSS23: An Audio-Visual Dataset of Spatial Recordings of Real Scenes with Spatiotemporal Annotations of Sound Events (2023 | NeurIPS 2023 Datasets and Benchmarks). Status: Peer-reviewed dataset paper. Reviewed: arXiv:2306.09126, revised November 2023. Access: Primary abstract and manuscript overview verified; no independent dataset-bias audit performed.

Research contribution. STARSS23 provides real multichannel audiovisual recordings with spatiotemporal annotations of sound events. Its contribution is shared evidence for studying detection and localization under real scene conditions. As a dataset and benchmark resource, it serves a different role from a separator, renderer, or generative model. [26]

Spatial9 assessment. We regard STARSS23 as an important anchor in real acoustic evidence. Its value does not depend on being a collection of commercial music masters. Real recordings and controlled scenes answer complementary questions, and neither automatically covers every production brief. Our assessment is that the dataset strengthens claims about perception in actual scenes while leaving music-specific fidelity and editing requirements to evidence suited to those tasks. Dataset relevance is a matter of fit, not a binary label. [26]

Our verdict. STARSS23 belongs in the evidence base for spatial understanding, alongside distinct resources for musical assets and authored scenes. We favor keeping recording provenance and scene-level separation intact when deriving questions or labels, because new annotations do not create new acoustic examples. For Spatial9, the architecture implication is a clear boundary between perception evidence and production acceptance. A model can be informative on real-world localization while a finished music workflow still has additional obligations to the approved performance and the creator's edits.


### SEP28: DCASE 2024 SELD distance baselines

Baseline Models and Evaluation of Sound Event Localization and Detection with Distance Estimation in DCASE2024 Challenge (2024 | DCASE 2024 Workshop, 23–25 October 2024). Status: Workshop proceedings paper. Reviewed: Published workshop PDF. Access: Full text: baseline definition, distance evaluation and conclusion.

Research contribution. The DCASE 2024 distance-estimation baseline paper expands sound-event localization and detection to include source distance, with associated model and evaluation definitions. It considers audio and audiovisual baselines and examines the interaction among detection, direction, and distance. The contribution makes a third spatial quantity explicit in a task often discussed mainly through angular accuracy. [27]

Spatial9 assessment. We regard the DCASE distance work as important because angular accuracy alone does not describe a three-dimensional scene. Its reported objective tradeoffs also caution against treating another input modality as an automatic improvement. Our production judgment distinguishes estimated physical distance from authored depth: a producer can use level, reverberation, and spectrum to create a presentation without committing to one recoverable distance in metres. Both quantities deserve clear names. [27]

Our verdict. Distance-aware perception is useful when the workflow needs physical scene information, while creative depth needs a separate control contract. We favor explicit units and meanings so an inferred distance cannot silently overwrite an artistic choice. For Spatial9, the architectural implication is to preserve the distinction between what the recording suggests and what the creator intends the listener to perceive. That makes depth controls more understandable and prevents an apparently precise number from carrying more physical authority than the task supports.


## Spatial representations and audio-language reasoning

We regard spatial perception and language reasoning as established, rapidly expanding research territory. The papers here make sound-location binding, spatial questions, motion, and varied capture formats accessible to language systems. That breadth makes a claim of merely combining an LLM with spatial audio insufficient as a distinctive position. Our emphasis is the contract between understanding and action: which source an instruction names, which reference frame it uses, when it applies, and what the creator can inspect or correct.

We favor retaining spatial evidence and source relationships through that contract. A fluent answer, an accurate acoustic estimate, and an appropriate edit are separate achievements with different failure causes. General-audio competence, input conventions, temporal continuity, and hardware assumptions also remain part of the architecture. For Spatial9, the strongest position is a clear and useful authoring system around these capabilities, with research credit attributed precisely and company judgments tied to concrete workflow requirements. Intelligence becomes valuable to a creator through understandable control, not through the vocabulary of the assistant alone.


### SEP16: LAION-CLAP

Large-scale Contrastive Language-Audio Pretraining with Feature Fusion and Keyword-to-Caption Augmentation (2023 | ICASSP 2023). Status: Peer-reviewed conference paper. Reviewed: arXiv:2211.06687, revised March 2024. Access: Abstract and metadata only.

Research contribution. LAION-CLAP learns a shared representation of audio and language through large-scale contrastive pretraining, feature fusion, and caption augmentation. It supports semantic tasks such as matching a recording to a description. That capability explains its frequent use as a component or evaluation signal in broader audio-generation and retrieval systems. [28]

Spatial9 assessment. We value LAION-CLAP as a semantic interface between sounds and descriptions. We do not accept a favorable semantic match as sufficient evidence that a spatial instruction has been followed. The desired sound can be present while its location or timing differs from the brief. Our judgment concerns the role assigned to the score; it does not assert that every checkpoint is wholly insensitive to spatial changes. Semantic correspondence remains a useful, bounded signal. [28]

Our verdict. CLAP belongs in semantic retrieval, routing, and assessment where its meaning is explicit. We favor a production interface that distinguishes 'the requested sound is present' from 'the requested spatial behavior is correct.' Those are separate claims even when one embedding responds to aspects of both. For Spatial9, the architecture implication is to resist turning a convenient similarity score into a universal acceptance criterion. A score is most useful when its label tells the creator which part of the brief it supports.


### SEP17: ELSA

Learning Spatially-Aware Language and Audio Embeddings (2024 | NeurIPS 2024). Status: Peer-reviewed conference paper. Reviewed: arXiv:2409.11369v2, 26 November 2024. Access: Full text methods, evaluation context, and explicit limitations inspected.

Research contribution. ELSA learns aligned representations of spatial audio and language, using spatially augmented recordings and captions. The paper evaluates semantic retrieval, spatial attributes, and captioning, and explores directional structure within its embedding space. It makes spatial information more accessible to language-based systems rather than treating audio solely as a monaural semantic signal. [29]

Spatial9 assessment. We regard ELSA as important prior work for giving language systems access to spatial meaning. The reviewed model leaves overlapping and moving sources to later work, so its contribution needs that version-specific boundary. We also distinguish directional structure in an embedding from a completed audio edit. A representation can support an interpretation without specifying which persistent source changes or how the rendered waveform realizes that change. Those are separate responsibilities in an authoring system. [29]

Our verdict. ELSA is a representation foundation, and its strongest architectural lesson is to make spatial meaning explicit before execution. We favor a language-to-scene interface that names the source, the reference frame, and the intended change independently of the embedding that helped interpret them. For Spatial9, this gives representation learning a clear place in the stack while preserving accountability at the action boundary. Understanding a direction becomes useful to a creator when it can inform an unambiguous, revisable instruction.


### SEP18: Spatial-CLAP

Spatial-CLAP: Learning Spatially-Aware audio--text Embeddings for Multi-Source Conditions (2026 | ICASSP 2026). Status: Peer-reviewed conference; official program verified. Reviewed: arXiv:2509.14785v1; ICASSP 2026 program metadata. Access: Full text methods and evaluation inspected, including simulated stereo conditions and folded direction classes.

Research contribution. Spatial-CLAP studies the binding between sound content and location in multi-source conditions. Its spatial contrastive objective introduces examples with swapped content-location assignments, requiring the representation to distinguish which sound occupies which position. The reviewed evaluation includes two-source training and unseen three-source mixtures. [30]

Spatial9 assessment. We regard Spatial-CLAP's emphasis on content-location binding as more relevant to editing than detecting the same sounds and positions independently. Its coarse horizontal classes and folded labels belong to the reviewed simulated two-microphone setting. They do not define the capabilities of all binaural systems. Our assessment is that the paper addresses an essential relationship, while the representation's tested granularity remains narrower than a continuous three-dimensional authoring contract. [30]

Our verdict. Correct source-to-location binding deserves an explicit place in a spatial production architecture. We favor identifying the object first and representing its authored coordinates separately, so a coarse perceptual category does not become an accidental limit on creative control. For Spatial9, the useful principle is relational correctness: knowing that a guitar and a drum are present is insufficient if their positions are exchanged. The paper advances that principle; an editing interface must carry it through to a stable source selection.


### NSEP01: BAT / Spatial-AST / SpatialSoundQA

BAT: Learning to Reason about Spatial Sounds with Large Language Models (2024 | ICML 2024, PMLR 235:61454–61469). Status: Peer-reviewed conference paper; later author revision reviewed. Reviewed: arXiv:2402.01591v3, 17 May 2025, plus ICML publication record. Access: Primary publication and full-text methods/limitations reviewed.

Research contribution. BAT combines the Spatial-AST acoustic encoder with a language model to answer questions about spatial sounds. Its SpatialSoundQA resource is built from simulated binaural scenes and covers event identity, direction, distance, and relationships. The paper is an important predecessor for systems that connect spatial perception with language reasoning; the encoder and dataset belong to the same research contribution. [31]

Spatial9 assessment. We regard BAT as essential prior art for connecting spatial perception with language reasoning. Its reviewed two-source, binaural, simulated setting and limited open-ended questioning define the scope of the evidence. The contribution is still substantial: spatial relationships become accessible through language. Our production assessment draws the next boundary at action. Answering where a sound is and deciding how a creator wants it changed are different responsibilities, even when they share a language interface. [31]

Our verdict. BAT is a foundational spatial-understanding system, and it rules out treating the basic perception-to-language idea as a new market claim. We favor positioning Spatial9 around the production contract surrounding such intelligence: explicit source references, inspectable instructions, and a clear connection between an accepted command and the authored scene. Research credit belongs to the demonstrated bridge; authoring value depends on what the complete workflow makes possible.


### SEP27: DSpAST

DSpAST: Disentangled Representations for Spatial Audio Reasoning with Large Language Models (2026 | ICASSP 2026). Status: Peer-reviewed conference; official program verified. Reviewed: arXiv:2509.13927v2, 1 November 2025; ICASSP 2026 program metadata. Access: Full text methods, evaluation definitions and baseline comparison notes inspected.

Research contribution. DSpAST separates semantic and spatial information within a spatial-audio encoder, building on Spatial-AST and evaluating the resulting representations inside BAT. The reviewed experiments use SpatialSoundQA and report improvements across detection, directional, distance, and reasoning tasks. The architecture targets the difficulty of asking one representation to serve several partly distinct acoustic objectives. [32]

Spatial9 assessment. We find DSpAST's factorization useful because semantic identity, direction, and distance need not benefit from an undifferentiated representation. Its reported improvements remain tied to the question definitions and reproduced baseline conditions. We regard that detail as essential to interpreting the result, not a reason to dismiss it. For a production assistant, the further distinction is between a correct answer about a scene and a valid instruction that changes the intended object. [32]

Our verdict. DSpAST supports an architecture that keeps spatial and semantic information inspectable before it reaches an action layer. We favor a scene representation with explicit attributes and source associations, so a mistaken answer does not disappear inside a plausible command. For Spatial9, the useful contribution is a better-defined perception interface, while execution needs its own requirements for identity, coordinates, and time. A language model's fluency is valuable when the information it acts on remains understandable and correctable.


### NSEP02: STARSS23 spatial QA baseline

Towards Spatial Audio Understanding via Question Answering (2025 | DCASE 2025 Workshop, 30–31 October 2025, Barcelona). Status: Published workshop proceedings paper. Reviewed: Published five-page proceedings PDF. Access: Full text task, dataset construction, evaluation and conclusion reviewed.

Research contribution. Towards Spatial Audio Understanding via Question Answering adds spatiotemporal descriptions and question-answer pairs to STARSS23 recordings. Its baseline uses first-order Ambisonic input and formulates answers as classification tasks covering events and spatial or temporal relationships. The study compares this scene-level QA supervision with a conventional frame-supervised SELD baseline. [33]

Spatial9 assessment. We value this work for grounding question answering in real spatial recordings and a defined answer space. Its classification formulation is narrower than unrestricted conversation, and its labels inherit existing annotation work. We regard that specificity as useful: a constrained answer can provide a clearer statement about a scene than an elaborate unsupported explanation. The production relevance is access to spatial and temporal relationships, with the scope of the answer kept visible to the creator. [33]

Our verdict. Structured spatial QA is a credible inspection layer for an audio workspace. We favor questions whose answers can refer to explicit events, intervals, and relationships in a scene record. For Spatial9, the architectural implication is to separate an explanation of the current scene from permission and instructions to change it. This paper supports the understanding side of that interface; a conversational production assistant also needs durable source identity and a defined action vocabulary rather than relying on answer fluency alone.


### NSEP03: Spatial-Omni / SO-Encoder

Spatial-Omni: Spatial Audio Understanding Integration in Multimodal LLMs via FOA Encoding (2026 | arXiv:2606.10738). Status: Preprint; peer-reviewed venue not independently verified. Reviewed: arXiv:2606.10738v2. Access: Full text architecture, benchmark scope and explicit limitations reviewed.

Research contribution. Spatial-Omni adds a dedicated first-order Ambisonic encoder alongside an existing audio pathway, converting spatial cues into tokens for multimodal language models. It introduces training and QA resources together with SO-Bench, which covers several spatial-understanding tasks. The reviewed second version evaluates both acquired spatial ability and retention of general audio capability. [34]

Spatial9 assessment. We regard Spatial-Omni's separate spatial pathway as a serious integration approach, with an explicit FOA input contract. Its twenty-second scope and reported general-audio degradation make the tradeoff concrete. Added spatial ability is valuable, but we do not treat it as a free expansion of a general assistant's competence. Our assessment favors preserving the coordinate convention and the distinction between semantic and spatial performance when interpreting the result. [34]

Our verdict. Spatial-Omni is a useful architecture for adding spatial perception to a broader language system when its representation assumptions match the workflow. We favor an explicit format boundary and distinct responsibilities for segment handling, source continuity, and action generation. For Spatial9, the product implication is that an assistant must remain dependable on the creator's general audio requests while gaining spatial capabilities. A specialized encoder earns its place through a coherent overall contract, not simply by adding another set of tokens to the model.


### NSEP04: MiDashengLM-Spatial / Spatial-Dasheng

MiDashengLM-Spatial: Unifying General Audio Understanding and Spatial Awareness (2026 | arXiv:2610.11156, first submitted 8 October 2026). Status: Preprint; peer-reviewed venue not independently verified. Reviewed: Primary indexed abstract dated 8 October 2026 and official author repository as retrieved 25 October 2026. Access: Primary indexed abstract and official repository reviewed; full manuscript fetch unavailable.

Research contribution. MiDashengLM-Spatial extends a general audio-language model with the Spatial-Dasheng encoder and hierarchical semantic-to-spatial conditioning. The authors describe a unified system for general audio understanding and spatial awareness, and provide an official repository with spatial and monaural benchmark summaries. Its October 2026 preprint is relevant to the rapidly developing intersection of audio foundation models and spatial perception. [35]

Spatial9 assessment. We regard the unification goal as strategically important: creators benefit from semantic and spatial understanding that work together. Our access here is limited to the primary abstract and official repository, so we take the described integration direction seriously without assigning a methods-level advantage or weakness. That is a firm evidence boundary. A benchmark summary alone is insufficient for us to rank the model's reliability across production inputs or to endorse a particular deployment configuration. [35]

Our verdict. MiDashengLM-Spatial belongs among the important integration efforts in this field, while the accessible record supports an architectural assessment rather than a detailed competitive ranking. We favor coordinated semantic and spatial representations that expose their input assumptions and preserve the creator's source references. For Spatial9, the market implication is clear: generic audio understanding and spatial awareness are converging research areas. A persuasive product position therefore needs a specific authoring contract and useful control, beyond announcing that both abilities exist in one model.


### NSEP05: The World is Not Mono / TWNM

The World is Not Mono: Enabling Spatial Understanding in Large Audio-Language Models (2026 | arXiv:2601.02954). Status: Preprint; peer-reviewed venue not verified through an accessible primary record. Reviewed: arXiv:2601.02954v1, 6 January 2026. Access: Full text conceptual framework and implemented methodological scope reviewed; OpenReview blocked by verification page.

Research contribution. The World is Not Mono combines semantic and spatial encoders for language-based understanding of binaural scenes. Its framework uses synthetic spatial data, representation fusion, and a training curriculum that includes reasoning-oriented optimization. The paper also proposes a broader hierarchy for auditory scene analysis, connecting objects, relationships, and higher-level interpretation. [36]

Spatial9 assessment. We regard the paper's focus on spatial-semantic binding as a necessary part of scene understanding. We also distinguish its broad conceptual hierarchy from the narrower modeled stage, where temporal context is not an explicit prediction target. That distinction preserves the value of the framework without treating every proposed capability as demonstrated. For production, our priority is the relationship between a named source and its attributes; a rich scene description is useful only if those associations remain coherent. [36]

Our verdict. This work reinforces scene relationships as established research ground and makes a purely conceptual novelty claim unpersuasive. We favor a production architecture that carries source identity, spatial attributes, and temporal scope into an explicit edit description. For Spatial9, the opportunity is the usefulness and clarity of that authoring interface: the creator needs to know which source changes, during which interval, and with what intended result. A broad reasoning vocabulary adds value when it connects to those concrete decisions.


### NSEP06: Dynamic Spatial Audio Spectrogram Transformer / DSAST

Spatial Audio Motion Understanding and Reasoning (2025 | arXiv:2509.14666). Status: Preprint; peer-reviewed venue not independently verified. Reviewed: arXiv:2509.14666v1, 18 September 2025. Access: Full text methods and motion-benchmark description reviewed.

Research contribution. Spatial Audio Motion Understanding and Reasoning introduces a dynamic spatial encoder with audio grounding to estimate events, directions, and distances over time. A language model receives structured spatial attributes to answer questions about moving sources. The paper also introduces a motion-oriented reasoning benchmark, extending the discussion beyond static spatial descriptions. [37]

Spatial9 assessment. We value DSAST's structured attributes because they create an inspectable boundary between motion perception and language reasoning. The architecture also carries a clear dependency: later answers rely on the earlier estimates. We treat that as a reason to retain those estimates, not as an unmeasured failure claim. The paper addresses observed movement explicitly, while the creative decision about how a source ought to move remains a different task. [37]

Our verdict. Structured motion understanding is a useful foundation for explaining a dynamic scene. We favor retaining time-resolved source attributes alongside any language interpretation so a creator can inspect the basis for a statement. For Spatial9, the architecture implication is a stable link between an observed event, its temporal extent, and the object an edit names. That link makes motion reasoning useful to authoring while preserving the distinction between describing an existing trajectory and choosing a musically appropriate new one.


### NSEP07: PhaseCoder

PhaseCoder: Microphone Geometry-Agnostic Spatial Audio Understanding for Multimodal LLMs (2026 | arXiv:2601.21124). Status: Preprint; peer-reviewed venue not independently verified. Reviewed: arXiv:2601.21124v2, 5 August 2026. Access: Full text architecture, selected evaluation and explicit limitations reviewed.

Research contribution. PhaseCoder takes multichannel audio and microphone coordinates to learn spatial representations that can transfer across array geometries. The paper evaluates localization and integration with a multimodal language model for reasoning and targeted transcription. This is relevant when spatial intelligence must work with varied capture hardware rather than one fixed recording arrangement. [38]

Spatial9 assessment. We regard PhaseCoder's use of microphone coordinates as an important step toward portability across capture arrangements. Its stated microphone assumptions and dominant-speech focus keep that portability bounded. Geometry awareness does not erase device acoustics or distance ambiguity. Our assessment is therefore positive about the input abstraction, while rejecting the stronger interpretation that any array or sound scene becomes equivalent merely because its coordinates are supplied. [38]

Our verdict. PhaseCoder is relevant to hardware-diverse spatial perception, with device and acoustic context retained as part of the input contract. We favor separating array geometry from calibration and separating perceived location from creative placement. For Spatial9, the architectural implication is an adaptable capture interface that records what is known about the microphones and what remains inferred about the scene. That supports a clearer path from varied recordings to usable scene information without treating geometric portability as universal acoustic invariance.


### NSEP08: SPUR

SPUR: A Plug-and-Play Framework for Integrating Spatial Audio Understanding and Reasoning into Large Audio-Language Models (2025 | arXiv:2511.06606). Status: Preprint; no peer-reviewed venue identified in the inspected primary record. Reviewed: arXiv:2511.06606v2, 13 November 2025. Access: Full text architecture, dataset construction, training, baseline distinctions and evaluation protocol reviewed.

Research contribution. SPUR adds a first-order Ambisonic adapter to existing audio-language models and introduces SPUR-Set for spatial reasoning. It combines spatial covariance features with learned adaptation while retaining the base model weights and training additional components. Evaluations integrate the approach into Audio Flamingo 3 and Qwen2.5-Omni. [39]

Spatial9 assessment. We regard SPUR's adapter approach as a practical way to connect spatial evidence with an existing language system. The paper's monaural probes are diagnostic controls, not a fair competitive ranking of spatial models. Its ten-second samples and LLM-judged open answers also define the reported evidence. We value the integration contribution while keeping those conditions visible; a gain over an input path without interchannel cues does not establish general production superiority. [39]

Our verdict. SPUR makes a persuasive architectural case for adding a spatial perception pathway rather than expecting language ability alone to recover missing acoustic evidence. We favor preserving channel conventions and source references through that pathway into a clearly defined action interface. For Spatial9, the implication is that spatial intelligence depends on what the system can actually hear as well as what it can say. An adapter can supply useful evidence; the authoring layer remains responsible for converting that evidence into understandable, controlled changes.


### NSEP09: AudioMotionBench / formerly AMPBench

AudioMotionBench: Evaluating Auditory Motion Perception in Audio LLMs (2025 | arXiv:2511.13273; revised title in January 2026). Status: Preprint; no peer-reviewed venue identified in the inspected primary record. Reviewed: arXiv:2511.13273v2, 22 January 2026; v1 was titled Spatial Blind Spot: Auditory Motion Perception Deficits in Audio LLMs. Access: Full text synthesis design, question formats, model evaluation, per-format results and explicit limitations reviewed; prior title verified.

Research contribution. AudioMotionBench evaluates motion understanding through controlled binaural scenes and language questions. The revised paper, originally titled Spatial Blind Spot, reports results close to chance on separate multiple-choice and binary tasks for the tested audio-language models. Its analytically rendered moving sources vary direction, pitch, speed, and noise to isolate particular cues. [40]

Spatial9 assessment. We regard AudioMotionBench as an important challenge to equating semantic fluency with motion perception. Its restricted synthesis and tested model set keep that conclusion bounded. The different chance baselines for multiple-choice and binary questions also make a single undifferentiated accuracy misleading. Our judgment is that the benchmark isolates a meaningful capability gap in its evaluated setting, while it does not justify declaring every audio model spatially blind or assigning one universal cause to failure. [40]

Our verdict. Motion understanding deserves its own evidence in any product that interprets moving sound. We favor an architecture that retains the spatial input and the time-dependent observations behind a language answer, because a plausible description cannot substitute for those signals. For Spatial9, the product implication is clear: motion-related assistance needs a stated perceptual basis and an explicit time reference. This benchmark's value lies in that diagnostic discipline, not in turning a provocative earlier title into a blanket verdict on the field.


### NSEP10: Movement-focused Spatial AQA with thinking and temporal masking

Spatial Audio Question Answering and Reasoning on Dynamic Source Movements (2026 | arXiv:2602.16334). Status: Preprint; no peer-reviewed venue identified in the inspected primary record. Reviewed: arXiv:2602.16334v1, 18 February 2026. Access: Full text architecture, synthesis boundaries, temporal masking method and question-specific results reviewed; compared with September 2025 DSAST manuscript for distinctness.

Research contribution. This paper jointly trains a spatial audio-language system to answer movement questions with an optional thinking mode. It studies how query-conditioned temporal masking interacts with reasoning. The approach differs from the authors' earlier DSAST framework, which supplies structured spatial estimates to an LLM rather than jointly training the complete audio-language path. [41]

Spatial9 assessment. We regard query-conditioned temporal focus as a useful way to study what evidence a movement question needs. The reviewed ten-second, front-hemisphere scenes and question-dependent benefits keep that usefulness specific. Temporal masking also has a precise meaning here: it selects intervals rather than isolating clean simultaneous stems. We therefore distinguish attention to relevant time from source extraction, and we do not treat a more elaborate reasoning response as proof that the underlying scene interpretation is correct. [41]

Our verdict. Selective listening is valuable when the selection retains the context needed by the question. We favor an explicit distinction between interval selection, source separation, and relational reasoning so one operation is not silently credited with another's capabilities. For Spatial9, the architecture implication is to preserve access to the wider scene when an instruction concerns several objects or their movement. Focus can make a request tractable; the production interface still needs to explain which evidence was selected and which source the resulting action names.


## Spatial generation, upmixing and moving scenes

We regard new-content generation, stereo upmixing, and revision of a persistent object scene as related production categories with different fidelity contracts. A creative brief defines which material may change and which must remain authoritative. The reviewed work includes moving sources, multichannel output, video-conditioned acoustics, and streaming; those contributions deserve accurate credit. Our assessment turns on what each method makes available to the creator and which forms of later control follow from its representation.

We favor a clear distinction between a generated presentation and an inspectable authoring state. Either can be the right deliverable, depending on the project. Source continuity, local revisions, playback format, and timing become architectural obligations when the brief requires them, rather than presumed failures of every generative method. For Spatial9, the product implication is to state the creative contract plainly and preserve it through the workflow. A convincing immersive result has value; continued control over that result is a separate value that the interface must make concrete.

For a gentle introduction to the sound-field representation used throughout this chapter, see our guide to Ambisonics explained.

[Image: The FOA convention made explicit: ACN order, SN3D normalization and a declared azimuth direction (code example 8).]


### SEP19: SpatialSonic

Both Ears Wide Open: Towards Language-Driven Spatial Audio Generation (2025 | ICLR 2025). Status: Peer-reviewed conference paper. Reviewed: arXiv:2410.10676v2, February 2025. Access: Full text: conditioning, datasets, evaluation and discussion.

Research contribution. SpatialSonic, presented in Both Ears Wide Open, generates spatial audio from language and visual conditions. Its training resources and spatial conditioning support multiple and moving sources. The contribution brings sound content and spatial behavior into a shared generation process, offering a different starting point from taking an approved recording and assigning its existing assets to a scene. [42]

Spatial9 assessment. We regard SpatialSonic as substantial evidence that generation can combine sound content with spatial behavior, including movement. Its creative value is strongest where the brief permits new material. The paper's visual-interpretation and format boundaries matter, but they do not erase that contribution. Our production distinction is about revision authority: a plausible newly generated scene and an edit to an already approved performance have different obligations to source identity and unchanged material. [42]

Our verdict. SpatialSonic is a credible approach to generating spatial content from a creative brief. We favor treating generation and asset-preserving authoring as distinct modes with explicit rules about what may change. A creator commissioning a new environment can welcome variation that would be unacceptable in an approved recording. For Spatial9, the architectural implication is to preserve that choice at the workflow level: generative freedom has value when it is intentional, and local revision requires a clear account of which assets remain authoritative.


### SEP20: ImmerseDiffusion

ImmerseDiffusion: A Generative Spatial Audio Latent Diffusion Model (2025 | ICASSP 2025). Status: Peer-reviewed conference paper. Reviewed: arXiv:2410.14945v2, February 2025. Access: Full text: conditioning, training, spatial evaluation and caveats.

Research contribution. ImmerseDiffusion generates first-order Ambisonic audio using a latent diffusion model with spatial and acoustic conditioning. The paper compares descriptive conditions with explicit parameters, making the distinction between language-level instructions and numerical spatial controls directly testable. It evaluates semantic and spatial behavior through several measures. [43]

Spatial9 assessment. We value ImmerseDiffusion for making descriptive and numerical spatial conditioning part of the same discussion. We also accept the authors' cautions about evaluator dependence and nonzero reference error. A spatial score is less informative when its representation is entangled with the conditioning path, and measured error needs an interpretable floor. Our assessment is that explicit controls improve the clarity of intent, while their presence alone does not settle perceptual quality or artistic suitability. [43]

Our verdict. Descriptive flexibility and numerical precision belong in a shared authoring design with a visible translation between them. We favor allowing a creator to express an intention in language and inspect the resulting spatial parameters before accepting a change. For Spatial9, this paper supports a distinction between the control that was requested and the acoustic behavior that was produced. The interface earns trust by keeping those meanings clear, rather than treating a well-formed parameter or a favorable embedding score as the finished result.


### Code example 8: Make the FOA channel convention explicit

Illustrative integration pattern. Related reviews: SEP22, R12, R13. Generative Ambisonics and neural spatial codecs exchange a structured sound-field representation. This small analytic encoder makes that representation visible: ACN channel order W, Y, Z, X; SN3D normalization; and azimuth measured positively toward the left from the front. [43, 113, 135]

```python
import numpy as np
def encode_foa(mono, azimuth_rad, elevation_rad):
    audio = np.asarray(mono, dtype=float)
    if audio.ndim != 1:
        raise ValueError("Expected one mono source")
    az = np.broadcast_to(azimuth_rad, audio.shape)
    el = np.broadcast_to(elevation_rad, audio.shape)
    x = np.cos(el) * np.cos(az)
    y = np.cos(el) * np.sin(az)
    z = np.sin(el)
    weights = np.stack([np.ones_like(az), y, z, x])
    return weights * audio[None, :]
```

Reading the example. The output has shape [4, samples]. Angles can be scalars or sample-aligned trajectories. The unity W gain and ordered directional components make channel mistakes observable before audio enters a neural model, a decoder or an export path.

Scope. This is a plane-wave encoding primitive, not an implementation of ImmerseDiffusion. It omits propagation delay, distance attenuation and reverberation. Models using a different normalization or channel order need an explicit conversion and their own sample rate.


### SEP21: Diff-SAGe

Diff-SAGe: End-to-End Spatial Audio Generation Using Diffusion Models (2025 | ICASSP 2025). Status: Peer-reviewed conference paper. Reviewed: arXiv:2410.11299v1. Access: Full text: representation, datasets, Table I and scope limitations.

Research contribution. Diff-SAGe generates first-order Ambisonic audio from a sound category and requested direction using a flow-based diffusion model. The reviewed experiments focus on static, non-overlapping, one-second events. Its results expose a useful tension: distributional generation metrics and estimated directional accuracy can favor different methods. [44]

Spatial9 assessment. We regard Diff-SAGe's controlled event setting as useful because it exposes a tradeoff between generation quality and directional estimates. The evaluated one-second, static, non-overlapping events define a focused task, not a complete production system. The estimator's nonzero reference error also limits how literally its numbers can be read. Our assessment gives the model credit for that bounded contribution without treating unstudied longer or moving scenes as demonstrated failures. [44]

Our verdict. Diff-SAGe is relevant as a spatial event-generation component, particularly where content category and direction are the intended controls. We favor assembling such components under an explicit scene contract that separately defines event duration, overlap, and later revisions. For Spatial9, the architectural lesson is to preserve several quality dimensions instead of rewarding a single score. A generated event can be convincing in content and imperfect in placement, and the creator needs those judgments presented clearly enough to decide what matters for the scene.


### SEP22: SonicMotion

Generating Moving 3D Soundscapes with Latent Diffusion Models (2026 | ICASSP 2026). Status: Peer-reviewed conference paper; IEEE record verified. Reviewed: arXiv:2507.07318v2, September 2025; v1 title was Dynamic Spatial Audio Soundscapes with Latent Diffusion Models. Access: Full text: moving-source parameterization, training, experiments.

Research contribution. SonicMotion extends latent spatial-audio generation to moving soundscapes using textual and numerical motion conditions. The reviewed training construction interpolates source angles between start and end conditions over short examples. It therefore addresses motion explicitly and should not be grouped with static-only generation systems. [45]

Spatial9 assessment. We regard SonicMotion as an explicitly dynamic model and give it credit for bringing movement into generation conditions. The reviewed start-to-end motion construction is meaningful control, while a production trajectory can require additional structure such as holds and local revisions. We distinguish the scope of that control from a failure claim. Generating audible movement and retaining an independently editable motion description solve related parts of the creative task. [45]

Our verdict. SonicMotion belongs in the conversation about moving soundscapes, with motion generation and motion authoring treated as separate product commitments. We favor representing a trajectory as an inspectable part of the scene when the creator needs to revise one interval while retaining the rest. For Spatial9, the architectural implication is temporal specificity: movement needs a source identity, a time reference, and an explicit path. The artistic decision about whether that movement serves the material remains distinct from the model's ability to generate it.


### SEP23: ImmersiveFlow

ImmersiveFlow: Stereo-to-7.1.4 Spatial Audio Generation with Flow Matching (2026 | arXiv:2601.12950; latest reviewed v2 dated 28 September 2026). Status: Preprint; named peer-reviewed venue not independently verified. Reviewed: arXiv:2601.12950v2; replaces v1 evidence for current comparison. Access: Full text: training split, downmix, objective and listening evaluation.

Research contribution. ImmersiveFlow generates twelve-channel 7.1.4 audio from stereo through conditional flow matching in a learned audio representation. The reviewed September 2026 revision uses 500 songs with a song-disjoint split and fixed-matrix stereo downmixes. It is a directly relevant learned upmixer, evaluated against commercial alternatives and through objective and listening measures. [46]

Spatial9 assessment. We regard the revised ImmersiveFlow as a directly relevant learned upmixer, and our assessment uses its 500-song, song-disjoint version. Earlier criticisms of the dataset or split no longer describe that record. Fixed-downmix pairs and short binauralized listening establish a particular evidence domain; they do not settle every mastering or playback context. We also distinguish multichannel output quality from the separate authoring question of whether individual sources remain editable. [46]

Our verdict. ImmersiveFlow deserves serious consideration when the deliverable is a generated 7.1.4 presentation from stereo. We favor judging that role on its stated input and output contract, then treating object editing as an additional workflow requirement when a client needs it. For Spatial9, the architecture implication is clarity about deliverables: a channel-based presentation and a persistent object scene provide different revision affordances. The relevant product choice depends on the creator's need for a finished render, continued control, or both.


### SEP24: SwanSphere

Towards Streaming Synchronized Spatial Audio Generation via Autoregressive Diffusion Transformer (2026 | arXiv:2605.30940, 29 May 2026). Status: Preprint; peer-reviewed venue not verified. Reviewed: arXiv:2605.30940v1. Access: Full text: streaming architecture, timing, dataset, limitations.

Research contribution. SwanSphere investigates streaming spatial-audio generation synchronized with video and text. Its architecture combines autoregressive planning with local diffusion and spatially informed training to produce first-order Ambisonic output incrementally. The paper brings synchronization and generation timing into the same research problem. [47]

Spatial9 assessment. We regard SwanSphere's incremental generation and synchronization as a serious architectural contribution. Its reported dense-scene and recording-configuration boundaries matter because a dominant event is not always the only production-relevant event. We also distinguish first-chunk responsiveness from the complete timing behavior of an interactive system. That distinction credits the streaming method while keeping capture, generation, buffering, and playback within the meaning of a user-facing latency claim. [47]

Our verdict. SwanSphere is relevant to spatial experiences in which audio must arrive while the audiovisual scene unfolds. We favor treating sustained delivery and source coverage as first-class design requirements alongside the initial response. For Spatial9, the architectural implication is that streaming cannot be added as a label to an otherwise unspecified pipeline: timing and scene state have to remain coherent across its components. The streaming design deserves its own product category, with synchronization and continuous operation central to the architecture decision.


### SEP25: DynFOA

DynFOA: Generating First-Order Ambisonics with Conditional Diffusion for Dynamic and Acoustically Complex 360-Degree Videos (2026 | arXiv:2602.06846; latest reviewed metadata v4, 23 July 2026). Status: Preprint; peer-reviewed venue not verified. Reviewed: arXiv:2602.06846v4, 23 July 2026. Access: Full text v4 methods, evaluation and explicit limitations inspected.

Research contribution. DynFOA conditions first-order Ambisonic generation on dynamic scene geometry and material-related information derived from 360-degree video. The reviewed version evaluates moving sources, multiple sources, and complex geometry, and models propagation cues such as occlusion, reflections, and reverberation. Its contribution connects visual reconstruction to spatial audio generation. [48]

Spatial9 assessment. We regard DynFOA's use of visual geometry and propagation cues as valuable evidence that scene appearance can inform spatial generation. The material lookup approximation and stated generalization boundaries also make the strength of that inference clear. A visually plausible acoustic result is not necessarily a uniquely recovered physical room. Our assessment credits the geometric conditioning while keeping inferred material properties distinct from directly established acoustic properties. [48]

Our verdict. DynFOA is a useful direction for visually guided spatial sound, with inferred acoustics requiring a clear status in the scene representation. We favor distinguishing measured, inferred, and artistically chosen parameters so a creator knows which values are evidence and which are interpretations. For Spatial9, the architectural implication is an acoustic scene model that can accommodate calibration where fidelity to a real environment matters. Visual conditioning can supply a coherent starting point without turning appearance into an unwarranted guarantee of physical equivalence.


## Binaural rendering and listener-dependent reproduction

We treat binaural rendering as a distinct layer of spatial audio production: it transforms an audio representation into the signals delivered to two ears. Its quality matters directly, but it does not determine which source belongs where or how an engineer revises a scene. Our assessment keeps those responsibilities explicit. A renderer can be excellent within a defined input–output contract, while the surrounding authoring system remains responsible for source identity, creative intent, and the meaning of the supplied spatial information.

We consider the information available to a renderer part of its technical identity. Source signals, pose, acoustic calibration, training domain, and listener assumptions all affect what its output means. BinauralGrad's acknowledged baseline constraint makes that principle concrete. We credit learned capture knowledge and measured acoustic responses as different sources of information; neither disappears from a fair interpretation. A result against an incompletely calibrated renderer cannot settle the relative merits of neural synthesis and conventional rendering in every production setting. [50]

We also distinguish binaural delivery from a static scene and processing throughput from interactive response. Two ear signals can represent changing spatial relationships. BinauralFlow's causal, streamable architecture deserves credit for addressing runtime behavior directly. Buffering, scheduling, pose updates, and device transport nevertheless remain part of the experience. We assess streaming as a substantive engineering property with an operating envelope, rather than as a universal quality label or a category that generative models are automatically excluded from. [53]

Our verdict across this chapter is that source fidelity, spatial perception, listener fit, and delay require separate technical judgments. For Spatial9, the product relevance is an architecture in which the intended scene and the means of reproducing it remain intelligible. Generic spatial speech, calibrated acoustic reconstruction, and interactive headphone rendering are legitimate but different briefs. We place each method according to the brief its evidence supports, and we regard the clarity of that placement as more useful than declaring one rendering family universally superior.

For the listener-side basics behind these papers, see headphone vs speaker spatial audio.

[Image: Binaural cues reach two ears; coordinate frames must agree before a neural renderer sees them (code examples 9 and 10).]


### B03: WarpNet / neural binaural speech renderer

Neural Synthesis of Binaural Speech from Mono Audio (2021 | ICLR 2021). Status: Peer-reviewed conference. Reviewed: ICLR 2021 author record and official repository; full paper unavailable. Access: Primary author abstract and official repository; OpenReview full PDF blocked.

Research contribution. Neural Synthesis of Binaural Speech from Mono Audio, often associated with WarpNet, combines neural time warping and learned rendering to transform speech using relative source and listener pose. The author abstract and official repository establish a speech-rendering task with spatial conditioning. They do not establish general music conversion or a complete production authoring system; the full paper could not be retrieved during this review. [49]

Spatial9 assessment. We take pose-conditioned neural rendering seriously as a way to connect intended geometry with ear signals. Two-channel delivery is compatible with a dynamic scene; channel count is the wrong basis for dismissing this work. Our assessment separates the source and trajectory supplied to a renderer from the acoustic transformation it performs. That distinction keeps source identity, listener assumptions, and creative placement intelligible within a production architecture. [49]

Our verdict. We place this paper in the rendering layer of the market, with particular relevance to spatial speech and interactive dialogue. Its contribution gives a clear reason to keep intended motion available upstream of the final headphone signal. For Spatial9, the important architectural conclusion is that rendering and authoring need clean boundaries: a useful neural renderer can serve a production system without being responsible for separation, mix decisions, or project-level editability.


### Code example 9: Reconcile coordinates before neural binaural rendering

Illustrative integration pattern. Related reviews: B06, B15. The released neural binaural speech dataset defines x forward, y right and z up. An editor using y left needs an explicit basis conversion. This fragment converts positions and rotation matrices together, preventing a silent left-right reversal at the model boundary. [49, 134]

```python
import numpy as np
# Example editor frame: x forward, y left, z up.
# Released dataset frame: x forward, y right, z up.
C = np.diag([1.0, -1.0, 1.0])
def to_dataset_frame(xyz_m, rotation_matrix):
    # Positions already share the receiver-centred origin.
    xyz = np.asarray(xyz_m, dtype=float)
    R = np.asarray(rotation_matrix, dtype=float)
    if xyz.shape[-1] != 3 or R.shape[-2:] != (3, 3):
        raise ValueError("Expected positions and 3x3 rotations")
    positions = xyz @ C.T
    rotations = C @ R @ C.T
    return positions, rotations
```

Reading the example. Positions are already receiver-relative; the two frames share an origin. Both the local and world bases of the supplied rotation are converted. Quaternion components cannot simply be renamed; decode and re-encode them using the documented convention. Geometry must retain its meaning across model boundaries.

Scope. The editor frame is illustrative. The fragment does not invoke WarpNet, validate rotation orthogonality or align the release's 120 Hz tracking to its 48 kHz audio. Those are separate input responsibilities.


### B06: BinauralGrad

BinauralGrad: A Two-Stage Conditional Diffusion Probabilistic Model for Binaural Audio Synthesis (2022 | NeurIPS 2022). Status: Peer-reviewed conference. Reviewed: NeurIPS 2022 proceedings paper. Access: Full methods and evaluation sections reviewed.

Research contribution. BinauralGrad uses two diffusion stages to model shared and ear-specific structure when synthesizing binaural audio from mono input and spatial conditioning. The full paper reports strong results on its speech task, but also explains that exact head-related and room impulse responses were unavailable for its DSP baseline. The authors acknowledge that this baseline performed worse than the corresponding baseline in the original dataset publication. That qualification materially affects comparisons with conventional rendering. [50]

Spatial9 assessment. We consider baseline information part of the result, not a footnote. A learned model that has absorbed a capture setup and a conventional renderer missing its calibration do not solve the same information problem. The reverse imbalance also matters when only a conventional method receives exact acoustics. Our assessment credits BinauralGrad's synthesis contribution while refusing to turn a comparison with unequal acoustic information into a general verdict on neural rendering versus DSP. [50]

Our verdict. We regard BinauralGrad as credible evidence for conditional diffusion in binaural speech synthesis. We do not read it as a sufficient reason to replace calibrated rendering as a default production choice. Its relevance to Spatial9 is the importance of matching the renderer to the available information and the required control. A compelling neural architecture earns a place in that decision through its specific operating contract, rather than through a blanket hierarchy of algorithms.


### B07: Neural Ambisonics-to-binaural renderer

Binaural Rendering of Ambisonic Signals by Neural Networks (2022 | arXiv 2211.02301). Status: Preprint; later venue not verified. Reviewed: arXiv:2211.02301v1, 4 November 2022. Access: Full primary HTML: recording setup, evaluation, tables, and subjective-test scope reviewed.

Research contribution. Binaural Rendering of Ambisonic Signals by Neural Networks learns a decoder from paired first-order Ambisonic and dummy-head recordings. The reviewed preprint uses 49 minutes of band recordings from one room, split into training and evaluation segments. It reports objective improvements and architecture-dependent listening results, with subjective scores collected on one test recording. The full text also shows that better objective scores from data augmentation can coincide with lower subjective scores. [51]

Spatial9 assessment. We treat decoding as an independent determinant of the listening experience. An Ambisonic intermediate does not guarantee convincing headphone reproduction, and waveform error alone cannot settle timbre or immersion. The disagreement between objective and subjective outcomes is especially useful here: it exposes the risk of optimizing a proxy as though it were the entire product. Our assessment keeps scene representation, decoder behavior, and listener experience separate. [51]

Our verdict. We regard this as a useful learned-decoder proof of concept with a narrow empirical base. The single-room material and limited listening excerpt constrain its authority as a general replacement for established decoders. For Spatial9, the lasting product lesson is renderer interchangeability: the value of a scene representation includes the ability to choose an appropriate decoder without confusing a change in reproduction with a change in the authored scene.


### B14: ZeroBAS

Zero-Shot Mono-to-Binaural Speech Synthesis (2025 | Interspeech 2025). Status: Peer-reviewed conference. Reviewed: arXiv:2412.08356v2 and Interspeech 2025 record. Access: Full methods and evaluation reviewed.

Research contribution. ZeroBAS combines geometry-based warping and scaling with a pretrained denoising vocoder to synthesize binaural speech without training on binaural recordings. Its full text explicitly distinguishes generalized acoustics from reproducing a particular room or listener head shape. That makes zero-shot transfer the meaningful contribution. It does not establish personalized HRTF reproduction or measured acoustic reconstruction of the listener's actual environment, and the paper should not be credited with those broader claims. [52]

Spatial9 assessment. We consider reduced dependence on specialized binaural training strategically important. It changes how a spatial speech method can be deployed across capture conditions. We also distinguish acoustic plausibility from reproduction of a particular room or listener. Our assessment is that a precise positional input does not justify an equally precise claim about every listener's perception. Generic acoustic assumptions have a legitimate role when their purpose remains explicit. [52]

Our verdict. We regard ZeroBAS as a practical contribution to generic speech spatialization with a clearly defined acoustic boundary. Its value to Spatial9's market is the possibility of useful spatial rendering without bespoke binaural training material. We place personalized reproduction and room reconstruction in separate categories. The product lesson is to align the promise with the information supplied: a plausible spatial speech experience and a calibrated acoustic replica are different deliverables.


### B15: BinauralFlow

BinauralFlow: A Causal and Streamable Approach for High-Quality Binaural Speech Synthesis with Flow Matching Models (2025 | ICML 2025). Status: Peer-reviewed conference. Reviewed: arXiv:2505.22865v1 and ICML 2025 proceedings record. Access: Full methods, streaming performance and capture appendices reviewed.

Research contribution. BinauralFlow makes causal, streamable binaural speech synthesis an explicit modeling objective, using flow matching and a streaming architecture. The full paper reports processing throughput on an RTX 4090 using approximately 0.683-second audio blocks; its human capture material uses one listener in one room. These details support the reported experimental configuration. Throughput faster than playback is valuable, but is not itself a measurement of complete interactive latency. [53]

Spatial9 assessment. We credit BinauralFlow for treating streaming as an architectural concern. We reject blanket claims that generative binaural models are inherently offline. We also separate processing throughput from the delay a person experiences: buffering, startup, pose updates, and scheduling each affect interaction. Our assessment is that a useful real-time claim describes the complete operating path, while listener fit and spatial fidelity remain separate dimensions of the product decision. [53]

Our verdict. We regard BinauralFlow as a serious causal speech-rendering reference. Its architectural direction matters for interactive applications, while the reported hardware and capture conditions bound the evidence available for broader deployment. For Spatial9, its relevance is the need to make runtime behavior part of renderer selection from the outset. Streaming capability is a substantive engineering property; it does not by itself determine acoustic generality, creative control, or the suitability of an entire production workflow.


## Visual evidence and spatial sound generation

We see visual information as evidence about a scene, with different roles in reconstruction and invention. A system can infer where supplied sound belongs, or it can create a plausible soundscape for images. Those tasks allow different changes to content, identity, and timing. Our assessment starts with that distinction. An approved soundtrack carries preservation obligations that a silent-video generation brief does not, while a faithful conversion says little about the ability to invent appropriate missing events.

We credit the lineage that already connects decomposition, visual inference, and spatial reconstruction. SpatialAudioGen and Sep-Stereo establish important precedents, while PseudoBinaural and geometry-

aware learning make supervision and scene assumptions explicit. For Spatial9, the authoring distinction is the meaning and persistence of the inferred elements: an editor needs to recognize which sound is being changed and why its location was chosen. A latent decomposition can support that goal without automatically constituting a complete editable scene. [54, 56, 57, 58]

We assign audiovisual agreement an editorial meaning. Visible dialogue, narration, music, off-screen action, and displaced effects do not share one compulsory placement rule. YT-SPEECH's visible-speaker curation makes that boundary concrete. Our position gives each sound's role priority over automatic visual alignment. A visible match is useful when it serves the intended scene; deliberate disagreement can be equally correct. Confidence in visual inference and authority to make a creative decision are different properties of a production system. [62]

Our verdict is that audiovisual spatial research is becoming more capable, while the product categories remain distinct. Short-form sound creation, reconstruction of visible speech, and conversion of an existing master carry different responsibilities. For Spatial9, the durable position is clear creative control over that relationship: what is inferred, what is supplied, and what the editor intends. We judge methods by the task they actually address and identify their relevance to production through source continuity, editorial roles, and the scope of the changes they permit.


### B24: SpatialAudioGen

Self-Supervised Generation of Spatial Audio for 360° Video (2018 | NeurIPS 2018). Status: Peer-reviewed conference. Reviewed: NeurIPS 2018 proceedings paper. Access: Full architecture and evaluation formulation reviewed.

Research contribution. SpatialAudioGen is an early and important precedent for reconstructing spatial sound from 360-degree video and mono audio. Its NeurIPS paper uses latent source separation and spatial localization to produce Ambisonic components. The formulation assumes a bounded set of simultaneous latent sources and an omnidirectional mono surrogate. It therefore already links decomposition with spatial reconstruction; later discussions of separating content and assigning spatial behavior should acknowledge this lineage rather than claim the combination originated recently. [54]

Spatial9 assessment. We consider decomposition plus spatial reconstruction an established research foundation, and we credit this paper's early role directly. Our authoring distinction concerns what happens to a latent component over time. A mathematically useful component need not automatically be a stable, named element an engineer can revise across an entire work. We assess production usefulness through the correspondence between those representations and the sources that carry an editor's intended changes. [54]

Our verdict. We regard SpatialAudioGen as foundational prior art for audiovisual spatial reconstruction. Its relevance to Spatial9 reaches beyond chronology: it prevents novelty claims based solely on separating content and assigning spatial behavior. Our verdict places the differentiating product work in the continuity and meaning of control. A useful authoring abstraction connects the inferred components to persistent creative decisions, while the underlying decomposition-and-spatialization idea retains the credit this research established.


### B01: 2.5D Visual Sound

2.5D Visual Sound (2019 | IEEE/CVF CVPR 2019). Status: Peer-reviewed conference. Reviewed: CVPR 2019 publisher record and author project; full paper unavailable. Access: Primary publisher abstract and author project/repository; full PDF access blocked.

Research contribution. 2.5D Visual Sound made an influential connection between visible sound sources and binaural reconstruction. Its task takes an existing mono soundtrack and corresponding video, then predicts two ear signals using visual spatial information. The publisher abstract and author project establish this contribution; the full paper was inaccessible in this review. It is therefore evidence for visually guided spatialization, without a basis here for detailed claims about its failure modes. [55]

Spatial9 assessment. We regard visible source location as valuable evidence for spatial reconstruction. We also draw a firm distinction between the location visible in a frame and the location a soundtrack needs. Narration, music, off-screen action, and deliberately displaced effects have their own editorial roles. Our position is that visual inference belongs inside an explicit creative brief: the picture informs placement, while the intended relationship between sound and image determines whether that placement is right. [55]

Our verdict. We consider 2.5D Visual Sound foundational to visually guided spatialization. Its product relevance is the translation of image evidence into an audible spatial hypothesis. For Spatial9, that hypothesis becomes useful authoring information when its source and editorial role remain identifiable. We do not equate automatic visual alignment with creative authority: the decisive production capability is retaining a clear distinction between an inferred location and an approved placement.


### B02: Sep-Stereo

Sep-Stereo: Visually Guided Stereophonic Audio Generation by Associating Source Separation (2020 | ECCV 2020). Status: Peer-reviewed conference. Reviewed: ECCV 2020 paper / arXiv 2007.09902 author PDF. Access: Full paper methods and training setup reviewed.

Research contribution. Sep-Stereo connects visually guided stereophonic generation with source separation through a shared learning framework. Mono audio and video supply the evidence, with MUSIC and FAIR-Play used in the study. The full methods describe a separation setting that does not model motion. This is a specific modeling boundary, rather than proof that all two-channel output must be static or that visual guidance cannot support moving sound. [56]

Spatial9 assessment. We value the link between decomposition and spatial presentation because errors in either can undermine the other. Our production standard adds persistent identity: an engineer needs the same voice or instrument to remain recognizable across entrances, edits, and scene changes. A separated component and an editable production object therefore carry different responsibilities. We credit Sep-Stereo's shared learning formulation while treating continuity of identity as a distinct authoring requirement. [56]

Our verdict. We regard Sep-Stereo as a strong precedent for separation-assisted spatialization. Its stated motion boundary limits the conclusions we draw about moving sequences, while leaving its central learning idea intact. For Spatial9, the strategic lesson is concrete: combining separation with spatial processing is established prior art. Differentiation rests in the precision of the resulting creative control, particularly the ability to associate a revision with the intended source throughout a work.


### B04: PseudoBinaural

Visually Informed Binaural Audio Generation without Binaural Audios (2021 | IEEE/CVF CVPR 2021). Status: Peer-reviewed conference. Reviewed: CVPR 2021 publisher record and author project; full paper unavailable. Access: Primary publisher abstract and author project.

Research contribution. PseudoBinaural addresses a practical obstacle: useful binaural training recordings are difficult to obtain at scale. The CVPR paper generates training supervision from mono audiovisual material, using synthesized spatial audio to teach visually informed binaural generation. Its publisher abstract and author project were accessible, but a full-text methods and limitations review was not. The contribution concerns obtaining supervision, rather than demonstrating that synthetic acoustics reproduce every real recording condition. [57]

Spatial9 assessment. We treat the construction of training examples as part of the model's technical identity. Synthetic supervision carries assumptions about space, acoustics, and audiovisual correspondence; those assumptions shape the problem the network learns. We reject the idea that synthetic data is inherently inferior, but we also reject treating its labels as measurements of an actual room. The distinction matters whenever an interface presents an inferred location with apparent physical precision. [57]

Our verdict. We regard PseudoBinaural as an important data strategy for spatial learning. Its value is reducing dependence on specialized recordings, and its production boundary is the difference between a plausible spatial prior and evidence about a particular scene. For Spatial9, this supports a clear product principle: generated supervision can inform inference, while the provenance of a spatial decision remains visible in how that decision is described and used.


### B05: Geometry-aware multitask binaural model

Geometry-Aware Multi-Task Learning for Binaural Audio Generation from Video (2021 | BMVC 2021). Status: Peer-reviewed conference. Reviewed: arXiv:2111.10882v1, 21 November 2021; BMVC 2021 paper. Access: Full primary HTML: methods, dataset splits, quantitative comparisons, and listener study reviewed.

Research contribution. This geometry-aware model adds spatial coherence, temporal geometric consistency, and room-response learning to video-guided binauralization. It introduces SimBinaural and evaluates on simulated environments and FAIR-Play, including held-out scene configurations and a listener study. The full text also distinguishes training splits when comparing PseudoBinaural. Its FAIR-Play configuration omits room-impulse-response prediction because that dataset lacks the required ground-truth responses, an important qualification when describing the complete framework. [58]

Spatial9 assessment. We favor explicit geometric reasoning over treating every spatial question as a generic audiovisual association. Geometry provides structure, but a visible room does not specify every acoustic material, hidden surface, or off-screen event. Our assessment therefore distinguishes geometric consistency from complete acoustic knowledge. This paper already addresses unseen scenes; its significance lies in making those constraints part of learning, rather than leaving spatial coherence entirely implicit. [58]

Our verdict. We see this as a substantive step toward scene-aware binaural generation. The changing availability of room-response supervision across datasets also makes the framework's information requirements central to its interpretation. For Spatial9, the product relevance is a disciplined separation of observed geometry, inferred acoustics, and chosen creative space. A geometry-aware model is most useful when those three meanings remain distinct instead of being collapsed into a single claim of scene accuracy.


### B11: CLUP

Cyclic Learning for Binaural Audio Generation and Localization (2024 | IEEE/CVF CVPR 2024). Status: Peer-reviewed conference. Reviewed: CVPR 2024 publisher abstract; full paper unavailable. Access: Primary publisher abstract; full PDF blocked.

Research contribution. CLUP connects binaural generation and sound-source localization through cyclic learning. The CVPR publisher abstract describes how these related tasks can provide mutual supervision, making spatial correspondence part of the learning objective. Full-paper access was blocked in this review, so the reliable description is limited to that contribution and its stated audiovisual scope. We cannot responsibly turn plausible production concerns into author-reported limitations or claim a particular failure rate. [59]

Spatial9 assessment. We value mutual constraints between spatial generation and localization because they give the learning process a coherent spatial objective. We also distinguish consistency with a visible source from consistency with an approved creative brief. Dubbing, narration, and displaced effects can require different audiovisual relationships. Our assessment is that cyclic agreement is useful technical structure, while editorial intent determines which correspondence a production actually needs to preserve. [59]

Our verdict. We regard CLUP as an important consistency mechanism for audiovisual spatial learning. Its relevance to Spatial9 lies in making spatial correspondence an explicit part of the system's reasoning. We do not assign one universal artistic meaning to that correspondence. A production architecture needs a clear account of each sound's role, so that a localization-informed decision supports dialogue, narration, or effects according to the intended relationship rather than an undifferentiated alignment rule.


### B16: OmniAudio

OmniAudio: Generating Spatial Audio from 360-Degree Video (2025 | ICML 2025). Status: Peer-reviewed conference. Reviewed: ICML 2025 proceedings abstract and official repository; full PDF unavailable. Access: Primary proceedings abstract and official repository.

Research contribution. OmniAudio studies generation of first-order Ambisonic sound from 360-degree video. It introduces the Sphere360 dataset, combines spatial and nonspatial pretraining, and uses panoramic and perspective visual branches. The official ICML proceedings abstract and repository establish the task and report objective and subjective results on that dataset. The full proceedings PDF could not be retrieved here, so detailed methodological limitations should not be attributed to the authors from this review. [60]

Spatial9 assessment. We distinguish creating sound for silent video from spatially converting an approved soundtrack. A plausible invented soundscape can satisfy the first brief while a changed voice, event, or timing violates the second. Visual understanding is relevant to both, but preservation has a different meaning in each. Our assessment places the user's supplied material and intended creative freedom at the center of the comparison, ahead of the shared spatial output format. [60]

Our verdict. We regard OmniAudio as an important contribution to panoramic scene generation. Its product relevance to Spatial9 is in the sound-creation branch of immersive production, where semantic fit and audiovisual relationships are central. We keep that role distinct from conversion of an existing master. The market gains clarity when a system states whether it is inventing missing sound, reconstructing spatial evidence, or transforming supplied audio; a common Ambisonic format does not erase those different purposes.


### B18: ViSAudio

ViSAudio: End-to-End Video-Driven Binaural Spatial Audio Generation (2025 | arXiv 2512.03036). Status: Preprint; later conference not verified. Reviewed: arXiv:2512.03036v1, 2 December 2025. Access: Full methods, dataset description and limitations reviewed.

Research contribution. ViSAudio generates binaural audio from perspective video, optionally conditioned on text, through a dual-branch flow-based approach. The reviewed preprint introduces BiAudio and discusses an eight-second clip scope, noise-related artifacts, and future multichannel or first-order Ambisonic output. It is a video-driven sound-generation system. Its reported limitations should be retained, while avoiding the stronger and unsupported inference that two-channel generation cannot represent movement or that every example necessarily exhibits artifacts. [61]

Spatial9 assessment. We see the combination of video and optional text as a useful route to sound creation guided by a scene. For production, continuity adds a separate responsibility: the same source needs a coherent identity and role across an edit. We distinguish that long-form requirement from the quality of an individual generated clip. We also keep supplied-audio preservation separate from silent-video generation, because the allowed creative changes differ fundamentally between those briefs. [61]

Our verdict. We place ViSAudio in short-form audiovisual sound creation. Its clip scope and stated artifacts matter when describing the maturity of that use case, without negating the ambition of the generation method. For Spatial9, the product relevance is the boundary between a generated asset and a maintained sequence. A clip can be valuable material for an editor even when the surrounding system must separately manage continuity, source identity, and changes across the work.


### B21: Localizer–Renderer / YT-SPEECH

Visually-Guided Spatial Audio Generation for 360° In-the-Wild Speech Scenes (2026 | Interspeech 2026). Status: Accepted conference per author arXiv record. Reviewed: arXiv:2608.24579v1, 25 August 2026. Access: Full methods, curation and evaluation reviewed.

Research contribution. The YT-SPEECH Localizer–Renderer work uses 360-degree video and the omnidirectional W channel to reconstruct directional first-order Ambisonic channels, retaining W. Its full text describes confidence-gated audiovisual localization and complex-domain rendering. Dataset curation excludes voice-over and nonvisible speakers, and the listening study uses nine participants with five clips per method. Those details make the visible-speech setting explicit and constrain broader claims about arbitrary in-the-wild soundtracks. [62]

Spatial9 assessment. We value the explicit relationship between localization confidence, rendering, and retention of the omnidirectional signal. We also treat sound role as an essential editorial variable. On-screen dialogue, off-screen speech, narration, music, and ambience can require different relationships to the image. Our assessment credits the visible-speech formulation without importing its alignment policy into every soundtrack. Keeping source content and the reason for a placement distinguishable is central to coherent authoring. [62]

Our verdict. We regard this as a focused visible-speech spatialization system with a meaningful preservation choice. Its curation boundaries make it an unsuitable basis for a universal claim about arbitrary soundtracks, while clarifying the setting in which its contribution matters. For Spatial9, the product lesson is role-aware spatialization: confidence in a visual match has value only within the intended editorial relationship. Retaining an input component and deciding where a source belongs are related but separate responsibilities.


### B22: PLACE

PLACE: Positional Latent Adaptation via Conditioned Embeddings for Binaural Audio Generation (2026 | arXiv 2610.00630). Status: Preprint. Reviewed: arXiv:2610.00630v1, 30 September 2026. Access: Full architecture, ablations and listener tables reviewed.

Research contribution. PLACE adapts AudioX for binaural generation using positional conditioning, latent adaptation, and interaural objectives across text, video, and optional audio inputs. The reviewed preprint's listener tables are task-dependent: the in-distribution text-to-audio average trails both comparison models, while video and out-of-distribution results improve. The paper therefore warrants a differentiated reading. A broad claim of universal superiority would erase a meaningful difference already visible in its reported evidence. [63]

Spatial9 assessment. We assign model claims at the level of the task they actually address. PLACE's differentiated outcomes make that principle especially concrete. Positional conditioning is useful, but content identity, timing, audible quality, and spatial agreement remain separate production concerns. Our assessment treats a coordinate as a request to the model, not as proof that every aspect of the rendered result obeys that request or that unrelated content remains unchanged. [63]

Our verdict. We regard PLACE as a significant exploration of adaptable spatial conditioning with strengths that depend on the input task. Its reported video and out-of-distribution results merit attention, while the text-to-audio result prevents an undifferentiated superiority claim. For Spatial9, the relevant product question is the relationship between spatial intent and retained content. We value controllability that makes that relationship explicit, rather than collapsing several generation tasks into one headline quality ranking.


## Stereo upmixing and Ambisonic representations

We distinguish reconstruction, creative reinterpretation, and representation design throughout this chapter. Stereo upmixing, mono-to-stereo generation, and Ambisonic order expansion begin with different spatial evidence and produce different kinds of information. More output channels do not erase those differences. Our assessment asks what is retained, what is estimated, and what is deliberately authored. That account is essential when a precise coordinate or a higher-order format appears to imply knowledge that the input did not uniquely provide.

We regard source separation and ambience extraction as established ingredients in surround conversion. The AES and IEEE modular approaches make a novelty claim based only on decomposition untenable. Our product distinction concerns the relationship between those components and an engineer's intentions. A vocals category can help organize a mix while still containing several performers; a residual category can span unrelated events. Category-based redistribution and precise authorship of an individual source therefore remain different levels of control. [64, 65]

We interpret representation claims according to the information they encode. Horizontal Ambisonics can be an intentional design scope, and inferred higher-order components address a different problem from encoding known source positions. Ambisonizer and generative upscaling illustrate those separate boundaries. We likewise value stereo consistency without treating it as proof of recovered three-

dimensional geometry. A convincing creative arrangement can differ from an unknown original scene, and an accurate return to stereo does not resolve every spatial ambiguity. [67, 70]

Our verdict is that preservation and useful revision are the most consequential production requirements across these methods. For Spatial9, the market position rests on a clear relationship between the approved material and the intended change. Source character, transients, ambience, spatial behavior, and delivery format each matter, and one cannot silently substitute for another. We credit reconstruction methods for what they retain and infer, and generative methods for the creative scope they offer, while treating export format as a delivery property rather than a quality claim.


### B09: Voice extraction plus primary-ambience upmixer

Deep Learning Based Voice Extraction and Primary-Ambience Decomposition for Stereo to Surround Upmixing (2023 | AES 154th Convention Express Paper 62). Status: Synopsis peer-reviewed; full express manuscript explicitly not peer-reviewed. Reviewed: AES 154th Convention Express Paper 62 opening and publisher metadata. Access: Primary paper opening and publisher metadata.

Research contribution. This AES express paper combines deep voice extraction with primary–ambience decomposition for stereo-to-surround upmixing. Its accessible opening and publisher record establish a five-channel conversion approach. The publication status requires precision: the synopsis was peer reviewed, while the full express manuscript is explicitly not peer reviewed. We did not access enough of the complete paper to attribute detailed artifacts, quantitative limitations, or generalization failures to its implementation. [64]

Spatial9 assessment. We see voice extraction and primary–ambience decomposition as a sensible way to assign different spatial treatment to different musical roles. We also distinguish that organization from a complete editorial understanding of a song. Backing vocals, doubled leads, and reverberation tails complicate the relationship between a broad category and an engineer's intended element. This is a production distinction, not an allegation of an artifact in the limited-access implementation. [64]

Our verdict. We regard this paper as a practical precedent for decomposition-led surround conversion. Its relevance to Spatial9 is the need to make category-level redistribution understandable without presenting it as fine-grained source authorship. A five-channel result can serve an important delivery brief; the separate authoring question is how precisely a musical intention maps to the material being changed. We place the conceptual value in that decomposition, without assigning unverified sonic strengths or defects.


### B25: Negru stereo-to-5.1 modular upmixer

Automatic Audio Upmixing Based on Source Separation and Ambient Extraction Algorithms (2023 | IEEE SpeD 2023). Status: Peer-reviewed conference. Reviewed: IEEE SpeD 2023 abstract and metadata; full paper unavailable. Access: Primary IEEE abstract and metadata.

Research contribution. This IEEE SpeD paper describes automatic stereo-to-5.1 upmixing through source separation and ambient extraction. The accessible IEEE abstract identifies drums, bass, vocals, and other as the separated categories. Detailed full-text methods and limitations were not accessible, so this review does not attribute specific quality defects or benchmark outcomes. The verified contribution is a modular conversion approach that combines content decomposition with distribution into a surround presentation. [65]

Spatial9 assessment. We regard broad source categories as useful organizational tools with a different granularity from editorial objects. A vocals category can contain multiple performers, while a residual category can contain several unrelated events. That distinction follows from the authoring problem and is not a measured failure of this implementation. Our assessment credits the modular approach while locating finer creative control in the relationship between category outputs and the particular elements an engineer intends to revise. [65]

Our verdict. We place this paper among the practical precedents for stereo-to-surround conversion. Its relevance to Spatial9 is the clear separation between category-based redistribution and detailed source authorship. We do not derive a sonic quality verdict from the limited-access record. The architectural verdict is nevertheless direct: named stem categories are a useful starting organization, and they do not by themselves establish the identity, persistence, or independence of every editable sound in a production.


### B10: MuCh-Res-U-Net spatial parameter predictor

A Deep Learning Approach to the Prediction of Time-Frequency Spatial Parameters for Use in Stereo Upmixing (2024 | DAFx 2024). Status: Conference paper. Reviewed: DAFx 2024 proceedings paper. Access: Full methods, results and discussion reviewed.

Research contribution. The MuCh-Res-U-Net study predicts time–frequency spatial parameters from stereo, using a spherical directional grid for subsequent upmixing. Its full discussion is unusually useful: the model learned lateral position, while front–rear mapping was unsuccessful and elevation remained approximate. These are reported findings for the evaluated model and data, not a proof that every stereo conversion technique must exhibit the same pattern or that wider output formats supply recovered ground truth. [66]

Spatial9 assessment. We draw a firm distinction between recovering a recorded scene and creating a convincing new one. Stereo does not uniquely specify every possible three-dimensional arrangement, so precise output coordinates can conceal an ambiguous inference. Our position is that creative placement is valuable in its own right. It does not become more credible when described as recovered height or rearward information that the input may never have uniquely encoded. [66]

Our verdict. We regard the reported front–rear and elevation limitations as decisive boundaries on this model's reconstruction claim. The paper is valuable precisely because it makes those boundaries explicit. For Spatial9, its product relevance is the distinction between estimated location and authored placement. We favor spatial controls that communicate their purpose clearly: a deliberate new arrangement is an editorial decision, while a claim of recovered geometry carries a different evidential obligation.


### B12: Ambisonizer

Ambisonizer: Neural Upmixing as Spherical Harmonics Generation (2024 | arXiv 2405.13428). Status: Preprint; current record says under review. Reviewed: arXiv:2405.13428v1, 22 May 2024. Access: Full method, evaluation and limitations reviewed.

Research contribution. Ambisonizer frames neural mono or stereo upmixing as spherical-harmonic generation. The reviewed version produces W, X, and Y components, excluding the vertical Z component. Its evaluation and limitations discuss stereo-decoded outputs that trail original stereo, reduced width, percussion modeling, and uncertainty in assessing the generated representation. These details define a horizontal Ambisonic experiment and its reported tradeoffs; they should remain visible when describing it as immersive audio research. [67]

Spatial9 assessment. We judge an intermediate representation by the transformations and audible behavior it supports. A horizontal Ambisonic design is a legitimate scope, and height becomes a separate requirement when the production calls for it. We reject channel count as a substitute for spatial quality. Order, dimensional coverage, image stability, and source fidelity describe different properties; treating them as interchangeable obscures the practical tradeoffs an engineer has to manage. [67]

Our verdict. We regard Ambisonizer as a useful horizontal upmixing experiment whose reported width and percussion tradeoffs directly matter to music production. Its contribution does not support a general claim of full three-dimensional reconstruction. For Spatial9, the important conclusion is that format generation must remain accountable to the source material and the intended scene. More spatial channels have little production value when the musical identity or intended presentation becomes less dependable.


### B17: DiffStereo

DiffStereo: End-to-End Mono-to-Stereo Audio Generation with Diffusion Transformer (2025 | Interspeech 2025). Status: Peer-reviewed conference. Reviewed: Interspeech 2025 proceedings paper. Access: Full methods and reported scope reviewed.

Research contribution. DiffStereo uses a diffusion transformer for end-to-end mono-to-stereo generation without requiring position annotations. The reviewed Interspeech paper demonstrates stereo output and describes its generation method and evaluations. It does not establish a named-object editing interface. That is a boundary of the demonstrated system, not evidence that its stereo signal lacks spatial information or that a generative prior has no value within a larger authoring workflow. [68]

Spatial9 assessment. We distinguish a useful spatial result from the precision with which an editor can revise it. Increasing overall width and moving one instrument while retaining the rest of a mix are different creative operations. Our assessment defines control through the relationship between the requested change and the affected content. Model conditioning and an appealing stereo image do not, on their own, establish that relationship at the level of a named source. [68]

Our verdict. We regard DiffStereo as a meaningful contribution to generative mono-to-stereo conversion. Its demonstrated scope places it in spatial result creation rather than a complete source-level authoring environment. For Spatial9, the relevant distinction is revision granularity: global spatial alternatives and localized production decisions have different product value. We credit the former on its own terms and reserve claims about the latter for an architecture that explicitly represents what the editor intends to change.


### B19: S2X-Net

S2X-Net: Stereo-to-Multichannel Neural Upmixing Network (2026 | IEEE Transactions on Audio, Speech and Language Processing 34:4362–4375). Status: Peer-reviewed journal; accepted author manuscript reviewed. Reviewed: 2026 author-accepted manuscript; IEEE publisher metadata. Access: Primary author-accepted manuscript hosted on ResearchGate; publisher direct access blocked.

Research contribution. S2X-Net is a close production-oriented comparator: it combines source separation and primary–ambience extraction for stereo-to-5.1 or 7.1 upmixing, with a downmix-consistent formulation. The accepted author manuscript describes positional object metadata as a future extension. This makes its channel-based conversion and round-trip behavior demonstrated contributions, while editable scene objects remain outside the implemented scope described in the paper. The journal record and author manuscript were reviewed together. [69]

Spatial9 assessment. We regard S2X-Net as direct prior art for production-oriented neural upmixing. Source separation followed by ambience handling and redistribution is an established technical direction; modularity alone is an insufficient differentiation claim. Our assessment focuses on the authoring loop around that conversion: identifying material, revising an intended element, auditioning the result, and delivering it. Downmix consistency is valuable within that loop, while remaining distinct from the correctness of a three-dimensional scene. [69]

Our verdict. We consider S2X-Net a central surround-conversion reference for Spatial9's market. Its downmix-consistent design addresses a concrete production concern, and the manuscript's treatment of positional objects as future work identifies a clear architectural boundary. Our verdict is that channel reconstruction and object authorship deserve separate claims. The opportunity in finer control comes from what an editor can intentionally change and retain, rather than from relabeling an existing separation-and-redistribution pipeline as novel.


### B20: Generative Ambisonic Upscaling / flow matching

Generative Learning for Ambisonic Upscaling (2026 | Accepted IEEE Transactions on Audio, Speech and Language Processing, 2026). Status: Journal acceptance verified in author/university records; full author preprint reviewed. Reviewed: arXiv:2609.23479v1, 20 September 2026; acceptance checked in author/university record. Access: Full formulation, evaluations and conclusion reviewed.

Research contribution. Generative Learning for Ambisonic Upscaling estimates missing higher-order components from lower-order Ambisonic speech while retaining the observed channels. The full author preprint evaluates generative approaches and identifies model size and computation as obstacles to immediate real-time use. Its perceptual comparisons should not be described as proving equality merely because some differences are nonsignificant. The meaningful task is conditional inference of missing spatial information, with explicit preservation of available input components. [70]

Spatial9 assessment. We distinguish estimation of missing spatial information from direct encoding of known sources. An upscaler reasons from an incomplete sound field; a renderer supplied with source signals and positions starts with a different information budget. Our assessment treats those as separate engineering tasks. The provenance of higher-order channels matters: measured content, analytically encoded content, and inferred content can share a format while carrying different meanings and different kinds of uncertainty. [70]

Our verdict. We regard this work as a substantial contribution to conditional Ambisonic restoration. Retaining observed channels gives the inference a clear anchor, while the added components remain estimates. For Spatial9, its relevance is the discipline of preserving available evidence while identifying what processing adds. We place the method in spatial reconstruction, with its reported computational burden relevant to deployment; the presence of higher-order output alone cannot establish recovery of the original scene.


### B23: Aura

Controllable Mono-to-Stereo Upmixing using Reward-tilted Flows (2026 | arXiv 2610.15967). Status: Preprint. Reviewed: arXiv:2610.15967v1, 11 October 2026. Access: Full formulation, evaluation and discussion reviewed.

Research contribution. Aura uses reward-tilted flows to control mono-to-stereo reconstruction through inference-time spatial rewards, without task-specific retraining of its prior. The reviewed preprint assumes a known forward degradation operator and reports listening judgments from five participants on eight excerpts. Its restoration results include a tradeoff between mid-signal fidelity and spatial or high-frequency improvements. These constraints matter when interpreting an appealing demonstration as a generally reliable conversion method. [71]

Spatial9 assessment. We value the separation of a generative prior from adjustable spatial objectives. It makes control an active part of reconstruction rather than a property fixed entirely during training. We also distinguish global interchannel statistics from the trajectory of a named source. Both are meaningful controls, but they operate at different levels. Our assessment interprets changes in fidelity according to the preservation brief, rather than treating every increase in spatiality as a free improvement. [71]

Our verdict. We regard Aura as a substantive contribution to controllable reconstruction, with the known degradation operator and fidelity tradeoff central to its scope. Its product relevance to Spatial9 is the explicit negotiation between spatial character and retained source content. We place reward-based spatial control alongside, rather than equate it with, source-level authorship. The decisive distinction is what the control governs: the overall reconstructed signal or a specific element whose identity and timing remain defined.


## Binaural enhancement and preservation of spatial cues

We regard enhancement as a spatial problem whenever processing changes the relationship between a desired source and its environment. Microphone-array rendering, target extraction, binaural denoising, and cue-aware objectives make that relationship explicit. These papers address a different product brief from generating or upmixing a complete soundtrack. Our assessment starts with the desired component: a clean speaker, a preserved environment, and an intact musical scene are distinct targets, each with a different definition of useful suppression.

We treat preservation as multi-dimensional. Cleaner individual ear signals can carry a changed interaural relationship, while better speech clarity can coincide with the loss of meaningful ambience. The cue-loss study makes disagreement between preservation measures concrete, and dynamic beamforming shows why spatial behavior matters independently of a large waveform-quality gain. Our position is that no single score has authority to erase those distinctions. The intended listening experience determines which relationships a processing stage needs to retain. [78, 77]

We also distinguish evidence about an algorithm from evidence about a deployed listening system. GCFSnet's real-time listener work adds an important perspective, while its transport and hardware assumptions remain part of the result. LBCCN's fixed-direction release illustrates a different operating contract. We credit compact processing and controlled studies within their defined scope. Neither a small model nor a real-time factor describes all of the synchronization, acoustic, and device conditions that determine a complete binaural product. [75, 76]

Our verdict is that the enhancement literature strengthens the technical meaning of fidelity across immersive audio. For Spatial9, its relevance is the clarity of the preservation policy: which source remains, which context matters, whose spatial perspective is represented, and which changes serve the creative purpose. We place these methods as specialized processing contributions with distinct operating envelopes. Their value lies in making those choices more explicit, rather than treating denoising, scene reconstruction, and creative authorship as interchangeable demonstrations of spatial intelligence.


### B08: MMP multichannel deep filtering

Model-matching Principle Applied to the Design of an Array-based All-neural Binaural Rendering System for Audio Telepresence (2023 | IEEE ICASSP 2023). Status: Peer-reviewed conference. Reviewed: arXiv:2210.11123 revised 6 March 2023; ICASSP 2023 author paper. Access: Full primary author PDF: method, moving-source simulation, listener study, and CPU evaluation reviewed.

Research contribution. The model-matching telepresence paper treats binaural reproduction, ambience preservation, and noise reduction as a joint microphone-array problem. Its learned multichannel deep-filtering approach is evaluated against model-matching and localization-based alternatives. The full paper includes a simulated moving-speaker scene, a 15-listener study, and CPU processing measurements. These are useful evidence for that capture-to-headphone configuration; they do not describe stereo-master upmixing or general creative sound design. [72]

Spatial9 assessment. We regard the definition of desired sound as the central design decision in enhancement. A background component can be interference in a conversation and essential context in a documentary. Preserving an environment and suppressing noise are therefore potentially competing objectives. Our assessment values the joint formulation because it makes that tension explicit, while placing editorial intent above the assumption that removing more background sound always improves an immersive experience. [72]

Our verdict. We place this work firmly in capture-to-headphone telepresence. It is relevant to Spatial9 as an example of spatial processing organized around a defined listening purpose. The product conclusion is that enhancement needs a preservation policy: target clarity, environmental character, and movement continuity have different roles. A complete production brief specifies which of those relationships matters, rather than treating a cleaner signal as an adequate description of the desired scene.


### B13: Array2BR lineage

End-to-end multi-channel speaker extraction and binaural speech synthesis (2025 | arXiv 2410.05739). Status: Preprint; later venue not verified. Reviewed: arXiv:2410.05739v2, 11 July 2025; revised Array2BR lineage. Access: Full primary HTML of revised v2: formulation, method, and simulated-room evaluation reviewed.

Research contribution. The latest reviewed version of the Array2BR lineage is titled End-to-end multi-channel speaker extraction and binaural speech synthesis. It combines extraction from noisy, reverberant microphone signals with binaural output and a magnitude-weighted interaural-level objective. The full version includes simulated-room evaluations of speech quality and spatial cues. The earlier Array2BR title and this revision represent one evolving paper, not two independent demonstrations of the same approach. [73]

Spatial9 assessment. We assess this work through the target it is designed to produce. A clean spatialized speaker is a coherent communication objective; an immersive music or documentary scene can require surrounding sources and reverberation to remain audible. Those briefs define error differently. We credit the joint extraction-and-rendering formulation while keeping removal of unwanted material distinct from preservation of a complete soundtrack and its environmental context. [73]

Our verdict. We regard the Array2BR lineage as a focused contribution to spatial speech extraction and reproduction. Its relevance to Spatial9 is the importance of an explicit target definition before enhancement enters a creative workflow. The product boundary is clear: isolating the intended speaker and retaining the intended scene are separate functions. A system's behavior is coherent only when the user-facing purpose matches which components the processing is designed to retain or suppress.


### NBA01: BCCTN

Binaural Speech Enhancement Using Deep Complex Convolutional Transformer Networks (2024 | IEEE ICASSP 2024, pp. 681–685). Status: Peer-reviewed conference. Reviewed: arXiv:2403.05393v1 and ICASSP 2024 conference record. Access: Full primary HTML: architecture, losses, matched and unmatched speech evaluation, and results reviewed.

Research contribution. BCCTN combines complex-valued convolutional processing and transformer attention for binaural speech enhancement, predicting ear-specific complex masks. Its ICASSP paper evaluates noise reduction, objective intelligibility, and interaural level and phase errors on simulated speech in isotropic noise, including reverberant conditions. The reported results involve tradeoffs: the strongest noise-reduction score does not always belong to the model with the strongest cue preservation. This is directly relevant to spatial audio processing beyond synthesis. [74]

Spatial9 assessment. We regard the relationship between the ears as part of source fidelity. Cleaner individual signals can still encode a different spatial scene, so independent channel quality is an incomplete definition of preservation. Our assessment also keeps desired ambience distinct from unwanted interference. A processing stage needs a clear account of which sounds and spatial relationships belong to the intended experience; a noise-reduction score cannot make that editorial decision on its own. [74]

Our verdict. We regard BCCTN as a relevant cue-aware enhancement reference. The reported tradeoff makes the technical verdict clear: stronger denoising and stronger spatial preservation cannot be treated as one interchangeable objective. For Spatial9, its product relevance is in the design of preservation criteria around processing, especially when clarity competes with environmental context. We place that contribution in enhancement; it does not transfer automatically into a claim about creative upmixing or individualized hearing benefit.


### NBA02: GCFSnet monaural and binaural enhancement

Real-time multichannel deep speech enhancement in hearing aids: Comparing monaural and binaural processing in complex acoustic scenarios (2024 | IEEE/ACM Transactions on Audio, Speech, and Language Processing, DOI 10.1109/TASLP.2024.3473315). Status: Peer-reviewed journal. Reviewed: arXiv:2405.01967v3, 30 October 2024; journal DOI verified. Access: Full primary author PDF: real-time listener experiment, objective evaluation, discussion, and limitations reviewed.

Research contribution. Westhausen and colleagues compare monaural and binaural GCFSnet speech enhancement with traditional hearing-aid processing in two acoustic scenes. The journal study includes objective metrics and real-time speech-in-noise testing with hearing-impaired listeners. Benefits depend on the scene, with binaural processing particularly useful for localized interferers. The authors also identify practical limits: binaural signals were available without a wireless-link delay, and the complete evaluation setup retained substantial hardware latency. [75]

Spatial9 assessment. We give this study particular weight for connecting model behavior to real-time use with listeners. We also keep its conclusions attached to the participant group, acoustic scenes, and hardware configuration actually described. Our assessment is that transport and device assumptions are part of a spatial audio system's technical identity. An algorithm's value cannot be separated from the path through which synchronized information reaches it and processed audio reaches the listener. [75]

Our verdict. We regard GCFSnet's deployment study as an important reference for evidence quality and system integration. Its scene-dependent findings and stated latency assumptions resist a simple claim that binaural processing is universally better or immediately ready for untethered devices. For Spatial9, the relevance is methodological and architectural: a credible product assessment connects the model to its actual listening context. We keep the hearing-aid conclusions within that specialized context rather than borrowing them as general benefits.


### NBA03: LBCCN

A Lightweight and Real-Time Binaural Speech Enhancement Model with Spatial Cues Preservation (2025 | IEEE ICASSP 2025, DOI 10.1109/ICASSP49660.2025.10888392). Status: Peer-reviewed conference. Reviewed: arXiv:2409.12444v1 and ICASSP 2025 author record; official repository caveat checked. Access: Full primary HTML: method, experiment, metrics, CPU cost, and official repository training-direction note reviewed.

Research contribution. LBCCN targets lightweight binaural enhancement with complex convolutional processing, selected low-frequency enhancement, and relative acoustic transfer-function estimation. The ICASSP work evaluates quality, intelligibility, spatial-cue errors, and CPU processing cost. Its official repository adds a material caveat: the released model is trained for a fixed target direction in diffuse noise, with a training instance targeting one direction. That release-specific condition should not be generalized into a claim that every possible LBCCN model has the same restriction. [76]

Spatial9 assessment. We value efficiency when its operating envelope is explicit. A compact directional enhancer can be an excellent design fit for a constrained brief, while moving-target behavior belongs to a different requirement. We assess the released checkpoint according to its documented conditions and avoid extending that restriction to every possible version of the architecture. Parameter count and processing factor are useful attributes, but they do not describe the entire device-level audio path. [76]

Our verdict. We regard LBCCN as a purposeful constrained enhancer whose release documentation is essential to understanding its value. For Spatial9, its product relevance is the discipline of matching a processing component to the scene it supports. A fixed-direction speech tool and a general spatial authoring system occupy different positions in the market. We credit the compact cue-preserving direction while keeping target selection, motion handling, and total system delay explicit in the product's operating contract.


### NBA04: Time-varying RTF-guided binaural beamforming

Interpretable Binaural Deep Beamforming Guided by Time-Varying Relative Transfer Function (2026 | arXiv 2511.10168, revised February 2026). Status: Preprint; peer-reviewed venue not verified. Reviewed: arXiv:2511.10168v2, 17 February 2026; v1 posted 13 November 2025 under an earlier title. Access: Full primary HTML: dynamic RTF tracking, distinct array/binaural setups, results, and target-reverberation assumption reviewed.

Research contribution. This preprint guides a neural beamformer with time-varying relative transfer functions, using tracked acoustic information to improve spatial behavior for a moving target. Its full text separates eight-microphone beampattern analysis from binaural cue evaluation and reports clearer spatial behavior despite only small signal-quality changes. The simulations assume negligible target reverberation; the binaural target is anechoic with reverberant babble noise. Dynamic tracking is therefore an existing contribution, not a missing feature to allege. [77]

Spatial9 assessment. We value inspectable spatial behavior alongside waveform quality. A beamformer can change how coherently it follows a source without a dramatic change in an aggregate quality measure. Our assessment focuses on the meaning of its guidance signal: an explicit acoustic relationship gives spatial control useful structure, while estimated guidance remains an estimate. The distinction between intended target identity and acoustic tracking is particularly important when a scene contains competing or changing sources. [77]

Our verdict. We regard time-varying transfer-function guidance as a strong architectural direction for spatially coherent enhancement. The negligible-target-reverberation assumption defines a material boundary on the reviewed evidence. For Spatial9, the product relevance is the separation of a target's identity, its acoustic description, and the processing that follows it. We favor that explicit structure over treating every improvement as a generic quality gain, while recognizing that dynamic tracking is already a contribution of this work.


### NBA05: Binaural reconstruction and joint cue-preservation losses

A Novel Binaural Cue Preservation Loss for DNN-Based Binaural Speech Enhancement (2026 | arXiv 2608.16299). Status: Preprint; peer-reviewed venue not verified. Reviewed: arXiv:2608.16299v1, 17 August 2026. Access: Full primary HTML: loss definitions, controlled evaluation, conflicting metric outcomes, and conclusion reviewed.

Research contribution. Amble and colleagues propose two losses for binaural speech enhancement: one penalizes masking-induced interchannel reconstruction error, while the other jointly represents level and phase cues. In controlled anechoic experiments, the losses improve different aspects of cue preservation, with the reconstruction-oriented loss still showing poor level-error performance. The authors explicitly leave reverberant evaluation to future work. The full paper therefore supports a nuanced tradeoff, rather than a claim that a single spatial metric settles preservation. [78]

Spatial9 assessment. We take the disagreement between cue measures as a direct challenge to single-score definitions of spatial fidelity. Optimizing one objective does not make other spatial properties disappear from the product brief. Our assessment separates speech quality, interaural relationships, perceived direction, and environmental context. A score closely related to the training loss describes one aspect of the system; it cannot stand in for all of the listening experience a product claims to preserve. [78]

Our verdict. We regard this preprint as particularly valuable for clarifying the limits of cue-preservation claims. Its contrasting loss outcomes support a multi-dimensional definition of fidelity, and the anechoic scope bounds its acoustic conclusions. For Spatial9, the product implication is direct: preservation needs an explicit account of which relationships remain intact. We reject a winner selected solely by the metric its objective favors, while crediting the paper for making the underlying tradeoff concrete.


### Code example 10: Expose a concrete binaural cue preservation invariant

Illustrative integration pattern. Related reviews: NBA02, NBA03. Cue-preservation claims need a stated invariant. Applying the same real mask to the two complex ear spectra leaves their ratio unchanged wherever the mask and denominator are nonzero. This compact baseline makes that limited preservation claim directly inspectable. [78, 75]

```python
import numpy as np
def shared_mask(stft_lr, mask):
    spectrum = np.asarray(stft_lr)
    gain = np.asarray(mask, dtype=float)
    if spectrum.ndim != 3 or spectrum.shape[0] != 2:
        raise ValueError("Expected [ear, frequency, frame]")
    if gain.shape != spectrum.shape[1:]:
        raise ValueError("Mask and STFT grids must match")
    gain = np.clip(gain, 0.0, 1.0)
    return spectrum * gain[None, :, :]
```

Reading the example. One gain is shared across ears at each time-frequency bin. That preserves the mixture's interaural level and phase relationship at that bin. It gives us an explicit reference against which a more selective enhancement method's cue changes can be described.

Scope. This is not the paper's loss or an enhancement model. Preserving the mixture ratio does not isolate the target or guarantee noise removal. Zeroed bins lose their ratio, and waveform reconstruction requires a compatible STFT and overlap-add procedure.


## Acoustic fields and room reconstruction

We distinguish three acoustic promises: suggesting a convincing room character, replaying a captured scene and reconstructing the behavior of a particular physical venue. These are all useful, but they require different information. A picture, a rough mesh and measured responses cannot be treated as interchangeable inputs. Our position is that an acoustic system earns clarity by stating which promise it makes and preserving the observations and assumptions behind its output.

We favor architectures that keep source intent, room state and rendering behavior identifiable. That makes a source move different from a geometry edit and an artistic reverb choice different from a claim about a measured room. The strongest contributions in this chapter give one of those relationships a useful representation. For our acoustic-twin direction, the decisive issue is the connection between a requested change and the acoustic mechanisms that actually respond to it.

[Image: Neural acoustic fields and DiffRIR model the paths that form a room impulse response; calibration defines what the model can claim (code examples 11 and 12).]

[Image: Acoustic digital twins model how sound travels and reflects in a specific space.]


### A01: NAF

Learning Neural Acoustic Fields (2022 | NeurIPS 2022 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: Published NeurIPS 2022 paper; author arXiv lineage 2204.00628. Access: Full text relevant methods and limitations inspected in the research review.

Research contribution. Learning Neural Acoustic Fields represents room propagation as a continuous function of emitter and listener location. A learned acoustic field supplies an impulse response that can be applied to different dry sounds. Its contribution is a reusable representation of a captured acoustic scene, rather than a separate prediction for every possible musical recording. The paper explicitly distinguishes generalization within one scene from generalization to multiple scenes. [79]

Spatial9 assessment. We value NAF because it makes the room a reusable part of the scene rather than tying every acoustic result to one recording. For us, that is an important architectural distinction: a creator needs to change the source without rebuilding the meaning of the space. The per-room boundary also matters. A fitted room representation is a useful asset, but its identity and calibration belong alongside the audio, not hidden inside an apparently universal effect.

Our verdict. Our preference is to retain an explicit relationship between the room model, its observations and the positions it supports. That structure makes a captured scene intelligible to an author and prevents a room change from being mistaken for a simple source edit. We regard NAF as foundational work for calibrated scene playback; its strongest product implication is a reusable acoustic scene asset with a defined scope, not automatic reconstruction of any venue.


### Code example 11: Keep coordinates explicit at the neural field interface

Illustrative integration pattern. We keep room identity, positions, orientation and spectral coordinates explicit. This original adapter contract illustrates the fixed-height, four-orientation SoundSpaces configuration in NAF; it is not an official inference API. [79, 141]

```python
import numpy as np
def query_field(adapter, src_m, dst_m, yaw_deg,
                freq_hz, time_s, bounds_m):
    # Original contract; adapter is bound to one room checkpoint.
    lo, hi = np.asarray(bounds_m, dtype=float)
    p = np.asarray([src_m, dst_m], dtype=float)  # floor XY
    if np.any(hi <= lo) or np.any((p < lo) | (p > hi)):
        raise ValueError("Outside calibrated room bounds")
    if yaw_deg not in (0, 90, 180, 270):
        raise ValueError("Orientation unsupported by this contract")
    xy = 2 * (p - lo) / (hi - lo) - 1
    ear, f, t = np.meshgrid([0, 1], freq_hz, time_s, indexing="ij")
    request = dict(source_xy=xy[0], listener_xy=xy[1],
                   yaw_deg=yaw_deg, ear=ear, frequency_hz=f, time_s=t)
    out = np.asarray(adapter(request))
    if out.shape != ear.shape + (2,):
        raise ValueError("Expected [ear, frequency, time, logmag/IF]")
    return out
```

Reading the example. The adapter must translate these normalized floor coordinates and physical frequency/time grids to the checkpoint's own conventions, including its learned embeddings. Its two output values are log magnitude and instantaneous-frequency phase representation. They are not two audio channels or a ready-to-convolve impulse response.

Scope. The adapter, weights and STFT reconstruction are deliberately outside this fragment. Bounds checks do not establish free-space occupancy or model accuracy. A different height, room, continuous orientation or Ambisonic output needs its own documented contract; this helper does not add those capabilities.


### A02: SoundSpaces

SoundSpaces: Audio-Visual Navigation in 3D Environments (2020 | ECCV 2020 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: ECCV 2020 primary abstract; historical sampling scope cross-checked against SoundSpaces 2.0. Access: Primary abstract and successor discussion; full original PDF blocked by content-size limit.

Research contribution. SoundSpaces introduced an audio-visual navigation benchmark built on simulated acoustics in reconstructed three-dimensional environments. An embodied agent receives acoustic and visual observations while learning to reach a sound source. The original system uses precomputed, discretely sampled acoustic responses; its successor explicitly addresses the restrictions this places on continuous positions. This record was reviewed through the primary abstract and the successor's description, rather than the inaccessible full original PDF. [80]

Spatial9 assessment. We see SoundSpaces as important because it connects hearing to action. In an interactive environment, a spatial cue has to remain useful as the listener moves and chooses a direction. That requirement adds something a polished stationary demo cannot reveal. We also keep navigation success separate from production quality: the acoustic information sufficient to guide an agent is not a complete specification for preserving the character of an approved soundtrack.

Our verdict. SoundSpaces earns its place in our architecture discussion as a reference for spatial perception and embodied behavior. Its value is the relationship between an acoustic scene and an agent's decisions, rather than a ready-made route from a stereo master to an editable mix. We regard it as a benchmark foundation for interactive applications, with scene behavior and creative authoring remaining separate responsibilities in a production system.


### A03: SoundSpaces 2.0

SoundSpaces 2.0: A Simulation Platform for Visual-Acoustic Learning (2022 | NeurIPS 2022 Datasets and Benchmarks peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: Published NeurIPS 2022 Datasets and Benchmarks paper. Access: Full text relevant methods, evaluation and limitations inspected.

Research contribution. SoundSpaces 2.0 extends simulation to continuously sampled source and microphone positions, with configurable geometry and acoustic materials. It supports visual-acoustic learning and evaluates the simulator against real observations. The authors discuss imperfect meshes, uncertain material coefficients and limitations of geometric acoustics. Those are explicit model and input boundaries, not evidence that the platform cannot produce useful training data. [81]

Spatial9 assessment. We welcome the move toward continuously configurable acoustic environments because source and listener positions are part of the scene, not merely a list of available clips. Our reservation concerns the authority assigned to the simulation. A visually convincing room can still contain acoustically significant geometry or material errors. We treat those inputs as part of the model's quality, rather than assuming a more capable renderer can compensate for an inaccurate description of the environment.

Our verdict. For us, SoundSpaces 2.0 is a strong infrastructure contribution for controlled scene construction and acoustic learning. Its practical role is clearest when a system retains the provenance of the geometry, materials and responses it produces. We favor it as a simulation layer with inspectable assumptions; a venue-specific acoustic twin also needs a connection to the physical room that the simulator alone cannot supply.


### A04: AViTAR

Visual Acoustic Matching (2022 | CVPR 2022 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: Author manuscript arXiv:2202.06875v2 and supplementary experiments. Access: Full text relevant methods, experiments and supplement inspected.

Research contribution. AViTAR, introduced in Visual Acoustic Matching, transforms source audio to match a target environment shown in an image. Its audio-visual transformer is trained through self-supervised acoustic alteration of web video. The main evaluation emphasizes indoor speech, while supplementary non-speech experiments also exist. It would therefore be inaccurate to dismiss the method as speech-only or to claim that it never tests broader audio material. [82]

Spatial9 assessment. We find visual acoustic matching compelling for editorial work because an image is often how a creator expresses the desired environment. That makes audiovisual fit a legitimate objective in its own right. We draw a firm distinction between an acoustic interpretation of an image and a measured account of a venue. Hidden construction and materials remain ambiguous even when the generated sound feels appropriate to the picture.

Our verdict. AViTAR is most persuasive to us as a way to translate visual intent into an acoustic treatment. In a production architecture, that treatment belongs in an editable layer so the creator can retain the visual relationship while changing its strength or character. We regard the paper as valuable prior art for visually guided sound design; its product role is acoustic matching, with venue reconstruction requiring a different input contract.


### A05: FAST-RIR

FAST-RIR: Fast neural diffuse room impulse response generator (2022 | IEEE ICASSP 2022 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: ICASSP 2022 author manuscript, arXiv:2110.04057. Access: Full text methods and experimental setup inspected.

Research contribution. FAST-RIR generates room impulse responses from room dimensions, source and receiver positions, and reverberation-time conditioning. The paper reports fast inference and evaluates acoustic parameters and far-field speech-recognition utility. Its training and principal experiments use bounded rectangular-room configurations and reverberation ranges. The demonstrated speed and downstream speech benefit are therefore tied to a stated simulation and application setting. [83]

Spatial9 assessment. We value FAST-RIR's emphasis on efficient response generation, especially where many acoustic variants are needed. Our concern is what the speed claim buys the creator. Generating a response quickly does not include obtaining reliable room parameters, deciding whether its coloration suits the material or managing the result through revisions. Those surrounding decisions determine whether an efficient model becomes a useful production component.

Our verdict. FAST-RIR fits most naturally into parameterized acoustic augmentation and controlled room-effect workflows. We see less justification for treating its bounded room model as a general representation of irregular venues. For our architectural priorities, efficiency is useful when the supported geometry and conditioning remain explicit. We regard FAST-RIR as a focused response generator with practical value inside its domain, rather than a sufficient foundation for an unconstrained acoustic digital twin.


### A06: MESH2IR

MESH2IR: Neural Acoustic Impulse Response Generator for Complex 3D Scenes (2022 | ACM Multimedia 2022 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: ACM Multimedia 2022 paper in author PDF arXiv:2205.09248. Access: Full text method, application evaluation and explicit limitations inspected.

Research contribution. MESH2IR encodes a scene mesh with a graph network and generates acoustic impulse responses, using energy-decay preprocessing during training. The authors report fast generation and applications to speech processing. Their conclusion explicitly identifies missing control over material absorption and scattering. It also notes that moving scene objects require renewed scene encoding, although source motion can be represented through a sequence of responses. [84]

Spatial9 assessment. We like MESH2IR's decision to encode the scene separately from individual response queries. That separation gives an acoustic system a reusable representation of its environment. The consequential boundary is between moving a source through a fixed room and editing the room itself. We regard that distinction as central to an authoring interface, because a source move and a structural edit have different implications for cached acoustic state.

Our verdict. MESH2IR is an appealing component for geometry-led playback when the environment is stable. Its limited material control makes it less directly suited to workflows in which changing surface treatment is itself the creative or engineering task. We favor the reusable scene-encoding idea while keeping material and geometry revisions explicit. We regard MESH2IR as a useful scene-conditioned generator whose adoption fit depends on how often the environment, rather than the source, changes.


### A07: AV-NeRF

AV-NeRF: Learning Neural Fields for Real-World Audio-Visual Scene Synthesis (2023 | NeurIPS 2023 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: NeurIPS 2023 primary abstract and author project description. Access: Primary abstract and detailed author project; full text not retrieved.

Research contribution. AV-NeRF learns audio-visual scene representations for synthesizing sound and imagery at novel viewpoints. Its author materials describe acoustic-aware generation informed by visual geometry and source-relative coordinates. This review verified the primary abstract and detailed project description; the full paper was not retrieved successfully. The supported contribution is novel-view reconstruction of recorded scenes, and detailed numerical criticism should remain limited by that access depth. [85]

Spatial9 assessment. We value AV-NeRF's integration of sound and imagery around viewpoint because an immersive scene needs to remain coherent as the observer changes position. That is a different responsibility from giving a producer independent control over every source. We do not read the absence of a documented editing workflow as a failure of scene reconstruction. We read it as the point at which a production architecture needs an additional representation of creative intent.

Our verdict. AV-NeRF is relevant to captured-scene experiences and spatial reconstruction, especially when the desired outcome is another view of the same environment. For authoring, we favor retaining source identity and edit decisions outside the reconstruction model wherever the application requires independent revision. For us, AV-NeRF is a valuable scene-synthesis direction that can sit beside an object editor; the documented contribution does not make the editor redundant.


### A08: NACF

Neural Acoustic Context Field: Rendering Realistic Room Impulse Response With Neural Fields (2023 | arXiv 2023 and AV4D workshop copy; journal acceptance not verified). Status: 2023 author preprint and AV4D workshop copy; journal acceptance not verified. Reviewed: arXiv:2309.15977v1, 27 September 2023; AV4D workshop copy. Access: Full author preprint architecture, SoundSpaces evaluation and ablations inspected.

Research contribution. NACF conditions an acoustic field on geometry, material properties and spatial information, adding temporal modeling and an energy-decay objective. The reviewed preprint evaluates room-response predictions on SoundSpaces and includes component ablations. It demonstrates that contextual information and suitable supervision can improve the stated response metrics. A journal acceptance was not independently established; this entry identifies the reviewed 2023 preprint and workshop lineage. [86]

Spatial9 assessment. We agree with NACF's premise that acoustic context deserves an explicit role in response prediction. Geometry and material information can make a model's behavior more understandable than a prediction detached from its environmental inputs. The cost is that context has to come from somewhere. We consider measured coefficients, inferred material labels and catalog defaults different kinds of evidence, even when they occupy the same numerical field in a model input.

Our verdict. NACF is attractive to us where useful environmental context is available and its provenance can remain attached to the scene. It is a weaker fit for a workflow that promises accurate acoustics from minimal inputs while concealing the assumptions used to fill the gaps. Our verdict favors contextual acoustic modeling as an architectural principle, with the source and uncertainty of each contextual input treated as part of the product rather than preprocessing trivia.


### A09: RAF

Real Acoustic Fields: An Audio-Visual Room Acoustics Dataset and Benchmark (2024 | CVPR 2024 peer-reviewed dataset/benchmark). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: arXiv:2403.18821v1, 27 March 2024; CVPR 2024. Access: Full author text dataset design, evaluation and explicit limitations inspected.

Research contribution. Real Acoustic Fields pairs densely measured room impulse responses with visual observations and tracked source and receiver poses. The dataset supports evaluation of acoustic-field models and transfer from synthetic data. The reviewed author paper explicitly states that the measurements cover one physical room in two configurations. Its contribution is therefore a valuable real-data benchmark with a bounded environmental scope, rather than broad evidence across many unrelated venues. [87]

Spatial9 assessment. We regard measured acoustic data as a necessary counterweight to the convenience of simulation. RAF matters because it lets a model confront a physical environment instead of only reproducing the assumptions of another algorithm. We also distinguish measurement density from environmental diversity. Many positions in one room can describe that room richly without representing the variation encountered across studios, theaters, homes and outdoor spaces.

Our verdict. For Spatial9, RAF is a reference resource for understanding the gap between a fitted acoustic field and measured propagation. It is particularly useful for diagnosing spatial interpolation within a captured environment. We do not turn that role into a claim about all venues. We regard RAF as a valuable measured benchmark whose depth strengthens room-specific analysis, while a broader acoustic-twin architecture still needs a wider account of environmental variation.


### A10: DiffRIR

Hearing Anything Anywhere (2024 | CVPR 2024 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: arXiv:2406.07532v1, 11 June 2024; CVPR 2024. Access: Full author text relevant modeling, residual assumptions, transformations and evaluation inspected.

Research contribution. DiffRIR, presented in Hearing Anything Anywhere, estimates interpretable acoustic properties from sparse measured responses and rough scene geometry. It models source directivity and reflections, and demonstrates source transformations and changed listening positions. Several propagation effects, including diffuse and higher-order behavior, are approximated through a spatially invariant residual. The method therefore provides a physical structure with deliberate approximations rather than a complete simulation of every acoustic mechanism. [88]

Spatial9 assessment. We find DiffRIR's interpretable structure especially relevant to creative control. When a source changes orientation or position, an author benefits from a model that expresses why the sound changes. The tradeoff is concentrated in the residual: a convenient approximation can conceal the portion of the acoustic scene that does not follow the same explicit rules. We therefore distinguish control over modeled mechanisms from control over the entire room response.

Our verdict. DiffRIR is one of the more directly relevant architectural references for an editable acoustic scene. We favor exposing source and propagation parameters while identifying the behavior handled by an approximation. That supports a clearer relationship between a requested edit and its acoustic consequence. We strongly favor DiffRIR's interpretability and controlled transformations, with the invariant residual setting a meaningful boundary on how broadly a fitted scene can be revised.


### Code example 12: Keep reflection parameters tied to acoustic paths

Illustrative integration pattern. For us, DiffRIR's useful control is the connection between a surface parameter and the paths touching it. These original NumPy fragments express its amplitude-reflection mapping and final temporal blend, following equations 3 and 5. [88, 142, 143]

```python
import numpy as np
def wall_amplitude(logits, band_hz, fft_hz):
    logits = np.asarray(logits, dtype=float)
    amp = np.exp(-0.5 * np.logaddexp(0.0, -logits))
    return np.interp(fft_hz, band_hz, amp)
def combine_paths(direct, reflected, src_ir, residual, gamma):
    # Path signals already contain propagation and reflection filters.
    paths = np.asarray(direct) + np.asarray(reflected)
    explicit = np.convolve(paths, src_ir, mode="full")
    residual, gamma = np.asarray(residual), np.asarray(gamma)
    if not (explicit.shape == residual.shape == gamma.shape):
        raise ValueError("Use one complete, aligned response length")
    if np.any((gamma < 0) | (gamma > 1)):
        raise ValueError("Blend weights must lie in [0, 1]")
    return gamma * explicit + (1 - gamma) * residual
```

Reading the example. The stable exponential computes sqrt(sigmoid(logits)): energy reflectivity becomes amplitude. Multiply a surface response into every affected path; an unobstructed direct path has no wall factors. Use increasing band frequencies and queries within their span. The fitted gamma varies over time.

Scope. Direct and reflected paths are modeled contributions, not waveform segments extracted by this code. Path tracing, minimum-phase reconstruction, propagation delay and attenuation are omitted. The residual approximates unmodeled effects; it is not a measured late-only stem. Editing geometry requires new paths and can invalidate the fitted residual.


### A11: DANF

Direction-Aware Neural Acoustic Fields for Few-Shot Interpolation of Ambisonic Impulse Responses (2025 | Interspeech 2025 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: Published Interspeech 2025 author PDF. Access: Full text methods, experiments and conclusion inspected.

Research contribution. DANF introduces direction-aware supervision for interpolation of Ambisonic impulse responses, including an intensity-vector objective and few-shot adaptation. Its experiments investigate ways to initialize and adapt the acoustic field to a room. The demonstrated representation is first-order Ambisonics. The paper's contribution is to joint spatial consistency across channels, rather than simply reducing independent waveform or decay errors for each channel. [89]

Spatial9 assessment. We consider DANF's emphasis on relationships between channels more consequential than a simple increase in output channel count. A sound field depends on coherent spatial structure; plausible channels considered individually can still describe an inconsistent direction. For us, the important architectural lesson is to preserve that joint meaning through prediction and rendering. Representation capacity, directional accuracy and the final headphone or loudspeaker experience remain distinct properties.

Our verdict. DANF is a focused contribution that makes directional structure part of learning rather than leaving it to a final decoder to repair. We favor that principle for acoustic fields and multichannel processing. Its first-order scope is a deliberate operating domain, not a reason to discount the result. Our position is that directional supervision belongs among the core design priorities of a spatial renderer; adding higher-order channels is valuable only when their information remains coherent.


### A12: PI-DANF

Physics-Informed Direction-Aware Neural Acoustic Fields (2025 | IEEE WASPAA 2025 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: WASPAA 2025 author report TR2025-142, 14 October 2025. Access: Full text methods, experimental scope and conclusion inspected.

Research contribution. PI-DANF adds physical relationships between Ambisonic pressure and directional components to a neural acoustic field. The reviewed experiments concern the first 100 milliseconds of first-order responses, sampled at 8 kHz, in ten simulated rectangular rooms. The authors explicitly leave complete responses in realistic environments for future work. These limits define a focused investigation of early directional structure rather than a finished full-band room renderer. [90]

Spatial9 assessment. We support the use of physical relationships when they constrain a meaningful part of the acoustic response. PI-DANF is useful precisely because its contribution is specific. The phrase physics-informed can otherwise imply more completeness than a model actually contains. For us, the early directional field, late reverberation and high-frequency character are separate architectural concerns, and a principled objective for one does not automatically settle the others.

Our verdict. We read PI-DANF as a component-level advance in directional consistency, with a narrower role than a complete room renderer. Its value to a production architecture is the discipline of assigning explicit physical constraints to the parts they describe. We favor the modeling choice for early directional structure; the reported system supplies a useful building block, while complete room behavior requires additional mechanisms.


### A13: AV-DAR

Differentiable Room Acoustic Rendering with Multi-View Vision Priors (2025 | IEEE/CVF ICCV 2025 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: arXiv:2504.21847v1, 30 April 2025; ICCV 2025 publication status. Access: Full author preprint architecture, capture requirements and evaluation inspected.

Research contribution. AV-DAR combines visual features with differentiable, physics-based acoustic rendering. The paper uses multi-view observations, beam tracing and learned reflection or residual components, and evaluates six real-world environments drawn from two datasets. Its capture protocol includes selected visual views and acoustic observations. The contribution is a hybrid method whose accuracy and data-efficiency results are evaluated under explicit input and training conditions. [91]

Spatial9 assessment. We favor AV-DAR's combination of physical structure and learned correction over treating those approaches as rival philosophies. A room model benefits from knowing which behavior is explained by geometry and which is inferred from observations. Our practical concern is total capture effort. A reduction in acoustic samples has a different meaning when it depends on additional imagery, reconstruction and preprocessing that the operator must also supply.

Our verdict. AV-DAR is a convincing architectural reference for capture workflows that already contain usable visual and acoustic information. We see its strength in combining complementary observations, rather than making capture disappear. In an acoustic-twin product, that favors an intake process that exposes what each observation contributes and what a scene update invalidates. We favor the hybrid design, with acquisition and maintenance costs central to its adoption suitability.


### A14: SOAF

SOAF: Scene Occlusion-aware Neural Acoustic Field (2026 | IEEE/CVF WACV 2026 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: Detailed claims from arXiv:2407.02264v1, 2 July 2024; WACV 2026 metadata verified separately. Access: Full 2024 author preprint methods and limitations; final WACV 2026 PDF blocked with HTTP403.

Research contribution. SOAF introduces scene-occlusion-aware acoustic modeling for novel-view audio-visual synthesis. The reviewed 2024 preprint combines global and local acoustic fields with directional attention. Its stated limitations include known source positions and the absence of explicit reflection and reverberation modeling. The work later appears in WACV 2026; the proceedings PDF could not be retrieved, so those detailed boundaries are attributed specifically to the reviewed preprint. [92]

Spatial9 assessment. We consider occlusion a distinct authoring concern because a wall or doorway changes the relationship between a source and a listener. SOAF's focus is useful to us for that reason: spatial credibility depends on more than a source coordinate. We also resist treating one learned environmental behavior as a complete room model. An architecture needs a clear account of how occlusion interacts with direct sound, reflections and reverberation.

Our verdict. Our judgment of the reviewed SOAF version is that it offers a focused approach to scene-dependent obstruction, best understood as one acoustic mechanism within a larger renderer. We favor keeping that mechanism identifiable so an edit to geometry has a clear consequence. Its practical relevance is strongest for spatial scene behavior; the documented preprint does not justify substituting its occlusion model for the full acoustic description of a venue.


### A15: Multimodal NAF

Multimodal Neural Acoustic Fields for Immersive Mixed Reality (2025 | IEEE Transactions on Visualization and Computer Graphics 31(5) peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: IEEE TVCG 31(5), 2025, author manuscript. Access: Full text introduction and explicit limitations inspected.

Research contribution. Multimodal Neural Acoustic Fields for Immersive Mixed Reality uses sparse audio-visual observations and poses to predict sound from unseen viewpoints. The paper combines learned audiovisual processing and reports immersive evaluation. Its explicit future directions include direct handling of 360-degree input, more explicit material-dependent attenuation and real-time rendering. The reviewed text therefore provides both a broad mixed-reality motivation and concrete boundaries for the demonstrated system. [93]

Spatial9 assessment. We find the mixed-reality framing useful because it puts perception and viewpoint into the same experience. A convincing predicted scene, however, is only one part of an interactive system. We treat observation coverage, preparation and runtime response as different product constraints. An author can accept preparation time for a captured environment while still expecting listener movement to receive an immediate and stable acoustic response.

Our verdict. Multimodal NAF is important to our view of audiovisual reconstruction, but its reported scope fits a reconstruction component more readily than a complete live headset audio stack. We favor an architecture that makes preparation and interaction separate responsibilities and preserves the distinction between observed and inferred scene content. For us, the paper advances immersive scene representation while leaving the runtime and panoramic-input requirements of a production experience as separate engineering commitments.


### A16: MiNAF

Explicit Context-Driven Neural Acoustic Modeling for High-Fidelity RIR Generation (2026 | Interspeech 2026 long paper, per author project and linked proceedings). Status: Interspeech 2026 publication identified by author project; detailed review uses author preprint. Reviewed: arXiv:2509.15210v1, 18 September 2025, plus author project identifying Interspeech 2026. Access: Author preprint method overview and detailed project; linked final proceedings PDF not retrieved.

Research contribution. MiNAF extracts explicit local geometric context from a rough room mesh to guide neural impulse-response prediction. Its distance-distribution features complement sparse acoustic observations, and the author project reports experiments with noisy or reconstructed geometry. The project identifies an Interspeech 2026 publication; detailed review here also uses the accessible September 2025 author preprint. This version distinction is retained rather than treating the earlier manuscript as an automatically identical final paper. [94]

Spatial9 assessment. We like MiNAF's use of explicit geometric context because it makes the information entering the acoustic model easier to identify. A rough mesh is a realistic input in many capture workflows, and tolerance to that input's imperfections has practical value. We do not equate geometric tolerance with tolerance to every capture error. Surface acoustics and source or receiver registration can remain wrong even when the room's broad shape is adequate.

Our verdict. MiNAF is relevant to an acoustic-twin architecture that has imperfect geometry and useful acoustic observations. We favor its explicit context over concealing all scene dependencies inside an undifferentiated representation. The product implication is to preserve separate accounts of shape, material and registration confidence. We regard MiNAF as a useful route for working with rough scene geometry, with the limits of each input remaining visible rather than being absorbed into a general claim of robustness.


### A17: FLAC

Few-shot Acoustic Synthesis with Multimodal Flow Matching (2026 | CVPR 2026 peer-reviewed; title verified in proceedings index). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: arXiv:2603.19176v1, 19 March 2026; CVPR 2026 proceedings listing. Access: Full author text conditioning, real-data transfer, evaluation and limitations inspected.

Research contribution. FLAC uses multimodal flow matching to generate plausible room impulse responses from sparse context. Its evaluation includes synthetic rooms, transfer to measured rooms and a listening study. The paper models uncertainty rather than assuming sparse observations determine a single response. It also reports limitations associated with simplified real-scene geometry and a latent representation whose adaptation was constrained by the small real-data collection. [95]

Spatial9 assessment. We welcome FLAC's recognition that sparse observations can support several plausible acoustic interpretations. That is a strength for creative exploration, where alternatives can be useful. It creates a different obligation for venue reconstruction, where the desired answer is tied to one physical environment. We treat variation among generated responses as a representation of possibilities, not automatic proof that the model knows the size or source of its errors.

Our verdict. FLAC fits our view of an acoustic suggestion system particularly well: it can express ambiguity instead of forcing one apparently certain answer. A production architecture also needs to retain the chosen response so an approved scene does not change unexpectedly during revision or playback. We favor this probabilistic approach to acoustic interpretation, with selection and reproducibility central to its product role and measured-room fidelity remaining a separate meaning of success.


### A18: Image2Reverb

Image2Reverb: Cross-Modal Reverb Impulse Response Synthesis (2021 | IEEE/CVF ICCV 2021 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: ICCV 2021 primary abstract and proceedings description. Access: Primary abstract/proceedings only; full paper requests unsuccessful.

Research contribution. Image2Reverb synthesizes a plausible reverberation impulse response from an environment image. Its stated application includes settings that are inaccessible or used artistically. The primary abstract and proceedings description support that visual-to-acoustic contribution; the full proceedings PDF was not successfully retrieved. The documented objective is plausibility conditioned on appearance, so exact recovery of a physically measured room should not be attributed to the paper. [96]

Spatial9 assessment. We see Image2Reverb as a useful expression of creative intent: a creator can point to a place and ask for an acoustic character associated with it. That is a meaningful interaction even when the image cannot uniquely determine the room response. We prefer to make that ambiguity part of the interface. A suggested acoustic treatment can be chosen for its artistic fit without being presented as a measurement of the pictured space.

Our verdict. Image2Reverb belongs naturally in visually guided sound design and room-effect selection. We favor keeping the resulting treatment editable and separate from the source so a producer can accept the character without surrendering control over the mix. We regard Image2Reverb as a strong conceptual fit for acoustic inspiration, with a clear product boundary: choosing a plausible room impression and reconstructing a particular venue are different commitments.


### NAC01: Listen2Scene

Listen2Scene: Interactive material-aware binaural sound propagation for reconstructed 3D scenes (2024 | IEEE VR 2024). Status: Peer-reviewed conference, accepted status verified from author primary sources. Reviewed: arXiv:2302.02809v4, 1 February 2024. Access: Full author text architecture, acoustic and perceptual evaluations, and explicit limitations inspected.

Research contribution. Listen2Scene generates binaural room responses from scene geometry, material information and source/listener positions. It combines graph encoding with conditional generation and evaluates acoustic measures and perceived plausibility. The reviewed final author version explicitly assumes static scene geometry and accurate semantic labels for material assignment, with training based on simulated responses. Listener movement through that static scene is supported, so static geometry must not be confused with an immobile listener. [97]

Spatial9 assessment. We regard Listen2Scene as an important connection between scene geometry and an interactive listening experience. It demonstrates why a room representation has to participate in playback rather than remain a visual asset. Our main architectural concern is the authority of the geometry and material assignments. A responsive acoustic model can reproduce those assumptions efficiently, but it cannot make an incorrect account of the environment physically accurate.

Our verdict. Listen2Scene is most relevant to captured environments whose geometry remains stable while the listener or source moves. We favor that separation of scene preparation and response generation, with scene edits explicitly distinguished from movement inside the scene. We value Listen2Scene as an acoustic propagation component; source identity, creative revision and delivery remain responsibilities of the surrounding authoring system rather than functions supplied by binaural response generation itself.


## Listener personalization and binaural filtering

We treat listener personalization as a distinct layer of spatial production. The creator defines the scene; the listener profile and playback system determine how that scene reaches the ears. Measurement, scanning, retrieval and generation offer different ways to acquire a profile, with different burdens for the user. We judge their suitability through that complete relationship rather than assuming a lower spectral error or faster reconstruction settles the listening experience.

Our architectural preference is to keep profile acquisition separate from continuous playback, and to make magnitude, timing and headphone assumptions explicit. We value uncertainty when it guides acquisition, compact representations when they simplify rendering, and dataset provenance when it prevents misleading personalization. These are direct judgments about the roles of the papers below. Their relevance to Spatial9 lies in the listener layer they can inform, not in treating every HRTF model as a competing authoring system.

[Image: Profile estimation and playback are separate stages, and a playback engine needs complete filters with timing (code example 13).]


### A19: HRTF Field

HRTF Field: Unifying Measured HRTF Magnitude Representation with Neural Fields (2023 | IEEE ICASSP 2023 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: Author PDF arXiv:2210.15196; ICASSP 2023. Access: Full text representation, interpolation/generation and explicit future work inspected.

Research contribution. HRTF Field provides a continuous neural representation of head-related transfer-function magnitude across spatial sampling schemes. It also learns a latent representation that supports interpolation and generation across subjects. The full paper explicitly identifies phase or direct impulse-response modeling as future work. Its contribution is therefore a useful representation of spectral filtering, not a claim to complete every component of a personalized binaural playback system. [98]

Spatial9 assessment. We value HRTF Field as a way to organize listener-dependent spectral information across sampling layouts. That representation can simplify how a renderer accesses a profile, but it does not choose the right profile for a person. We place selection, timing and headphone behavior beside spectral accuracy in our architecture discussion. Each can change the perceived result even when the stored magnitude response is represented faithfully.

Our verdict. HRTF Field is an appealing representation layer for interpolation and profile management. We favor a modular use in which magnitude, timing and playback compensation have explicit responsibilities, rather than assigning the entire listening experience to one neural field. We regard HRTF Field as a useful foundation for handling HRTF spectra; a personalized product still needs a coherent listener profile and rendering path around that representation.


### Code example 13: Preserve both ears timing when a profile reaches playback

Illustrative integration pattern. Related reviews: A20, A21. We separate profile estimation from playback. This static renderer accepts a complete left/right HRIR for one listener and direction. HRTF Field's magnitude output alone does not satisfy that interface; NIIRF also handles interaural timing separately. [98, 99, 144, 145, 146, 147]

```python
import numpy as np
def binaural_static(x, hrir, fs_x, fs_h):
    # hrir: [left/right, taps], with all required delays included.
    x = np.asarray(x, dtype=float)
    h = np.asarray(hrir, dtype=float)
    if fs_x != fs_h or fs_x <= 0:
        raise ValueError("Signal and HRIR sampling rates must match")
    if x.ndim != 1 or h.ndim != 2 or h.shape[0] != 2:
        raise ValueError("Expected mono signal and two-ear HRIR")
    if x.size == 0 or h.shape[1] == 0:
        raise ValueError("Signal and HRIR must be nonempty")
    n = x.size + h.shape[1] - 1
    X = np.fft.rfft(x, n=n)
    H = np.fft.rfft(h, n=n, axis=-1)
    return np.fft.irfft(H * X[None, :], n=n, axis=-1)
```

Reading the example. Padding to the full convolution length prevents circular wraparound and retains the decay. The complex spectra preserve the supplied filters' phase and interaural delay. We preserve their relative gains instead of normalizing each ear independently.

Scope. Apply any separate SOFA Data.Delay before calling this function. Magnitude estimates need a justified phase-and-delay reconstruction. This is an offline, fixed-direction convolution, not personalization, head tracking or a streaming engine. Filter transitions, headphone compensation and safe playback gain remain separate responsibilities.


### A20: NIIRF

NIIRF: Neural IIR Filter Field for HRTF Upsampling and Personalization (2024 | IEEE ICASSP 2024 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: ICASSP 2024 author report TR2024-026. Access: Full text methods and experiments inspected.

Research contribution. NIIRF predicts compact parametric IIR filters for HRTF upsampling and listener adaptation. The method uses a stable filter parameterization and treats interaural timing independently of the modeled filter response. These are explicit parts of its design. A critique should therefore examine how the components work together rather than incorrectly claiming that the authors ignored filter stability or that every timing cue must be represented inside the same network. [99]

Spatial9 assessment. We find NIIRF's compact, stable filter representation attractive because it connects personalization to an implementable signal-processing structure. Its explicit separation of interaural timing is also a sensible architectural choice. We do not require every cue to be learned inside one network. What matters to us is that the components form a consistent listener response and remain manageable as the source or head changes direction.

Our verdict. NIIRF is particularly relevant where filter representation and playback constraints matter alongside reconstruction accuracy. We favor its modular treatment of filtering and delay, with profile acquisition kept distinct from continuous rendering. That makes the architecture easier to reason about without assuming a particular device-level speed advantage. We favor NIIRF's compact, structured personalization because its practical value lies in an integrated filter-and-timing path, not parameter count alone.


### A21: RANF

Retrieval-Augmented Neural Field for HRTF Upsampling and Personalization (2025 | IEEE ICASSP 2025 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: ICASSP 2025 author report TR2025-029. Access: Full text methods, evaluation and conclusion inspected.

Research contribution. RANF combines sparse personal HRTF measurements with retrieval of acoustically similar subjects and neural-field adaptation. Its formulation estimates dense spectral information and interaural timing. The authors identify a dependency on requested directions being present in the retrieved subjects' dense data, with off-grid retrieval modeling left for future work. The method consequently joins data selection, personal measurement and adaptation rather than eliminating those stages. [100]

Spatial9 assessment. We take RANF seriously because it combines an individual's measurements with information from other listeners rather than treating personalization as a generic profile-selection problem. The tradeoff is operational: the retrieval collection, available directions and adaptation process become part of the product. We consider that dependency acceptable in a calibrated workflow, provided it is visible. It is less compatible with a promise of effortless personalization that obscures what the listener must supply.

Our verdict. RANF fits a measurement-assisted personalization service more naturally than a zero-input consumer feature. We favor retaining the origin and coverage of the retrieved data alongside the resulting listener profile, so missing directions and adaptation decisions remain understandable. We regard RANF as a substantive approach to using sparse personal evidence efficiently, with onboarding effort and retrieval coverage determining its architectural fit as much as reconstruction quality.


### A22: Spherical neural process

HRTF Interpolation using a Spherical Neural Process Meta-Learner (2024 | IEEE/ACM TASLP 32 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: arXiv:2310.13430v1, 20 October 2023; 2024 journal metadata. Access: Full author preprint task definition, spectrum/uncertainty evaluation and scope inspected.

Research contribution. The spherical neural-process meta-learner models HRTF interpolation on a sphere and supplies predictive uncertainty. Its reviewed experiment corrects time-aligned spectra from a population-mean starting estimate rather than implementing an entire personalized listening loop. The paper evaluates interpolation accuracy and uncertainty calibration. This is a focused contribution to data-efficient correction and to deciding where additional information might be most valuable. [101]

Spatial9 assessment. We consider uncertainty useful when it changes a calibration decision, such as where more information is worth acquiring. A confidence estimate that never affects the workflow has limited value to a listener. We also distinguish confidence in an interpolated spectrum from confidence in the complete perceptual result. Timing, playback compensation and individual perception can remain relevant even when the spectral prediction is internally consistent.

Our verdict. The spherical neural process offers a valuable architectural idea: a profile representation can describe both its prediction and its uncertainty. We favor using that information to organize calibration effort rather than hiding it behind a single average reconstruction score. We consider the spherical neural process especially relevant to adaptive profile acquisition, while its focused spectral-interpolation contribution remains one component of a complete personalized playback path.


### A23: GraphNF-SCA

Graph Neural Field with Spatial-Correlation Augmentation for HRTF Personalization (2026 | AAAI 2026 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: arXiv:2511.10697v1, 12 November 2025; AAAI 2026 proceedings status. Access: Full author text architecture, multiple-dataset experiments and ablations inspected.

Research contribution. GraphNF-SCA uses graph-based personalization and spatial-correlation modeling to predict HRTFs, followed by fine-tuning. The reviewed paper includes SONICOM evaluations and additional measured or simulated HUTUBS and CIPIC experiments. It reports spectral and interaural-level improvements under different measurement densities. It would therefore be inaccurate to suggest that the work never examines multiple datasets or that its evidence is confined to a single benchmark. [102]

Spatial9 assessment. We credit GraphNF-SCA for examining several datasets and measurement densities; its evidence deserves more weight than a critique based only on an assumed single benchmark. Our interest is how spatial relationships across a profile support personalization from limited observations. We also keep average reconstruction improvements separate from the experience of an individual listener. Atypical anatomy, capture quality and playback conditions can matter without invalidating the reported dataset-level contribution.

Our verdict. GraphNF-SCA is a serious reference for reconstruction methods that exploit structure across directions. We favor that use of correlation while retaining the individual's measurements and the profile's operating assumptions as explicit assets. Its cross-dataset work strengthens its relevance to our architecture discussion. We favor this structured personalization approach, with the creator's scene and the listener's rendering profile remaining separate layers rather than competing explanations for spatial quality.


### A24: SPARO

HRTF Upsampling and Personalization via Neural Operators (2026 | arXiv preprint; peer review not established). Status: Preprint or workshop lineage; journal/conference acceptance not established beyond stated source. Reviewed: arXiv:2610.20400v1, 14 October 2026. Access: Full text methods, reconstruction timing and dataset protocol inspected.

Research contribution. SPARO formulates HRTF upsampling and personalization with neural operators and reports reconstruction accuracy and adaptation timing. The reviewed October 2026 preprint models both magnitude and interaural delay. Its computational measurements concern listener-profile reconstruction on stated GPU hardware, and its dataset allocation differs from some published baseline settings. Peer-reviewed acceptance was not established for this record. [103]

Spatial9 assessment. We view SPARO's reconstruction and adaptation work as relevant to the preparation of a listener profile. That is a useful place to reduce effort, but it is distinct from shortening the delay of audio already playing through that profile. We also read its comparative results in the context of the stated data allocations. Changing the information available to a method changes what a speed or accuracy comparison can establish.

Our verdict. SPARO is most relevant to us as a profile-construction architecture, with magnitude and timing considered together. We favor keeping that preparation stage separate from the runtime renderer so a useful reconstruction method is not burdened with an unrelated playback claim. We regard neural operators as a promising personalization approach here; the published timing evidence supports discussion of profile generation, not a blanket claim about interactive audio latency.


### A25: S2RNF

HRTF Personalization via Sim-to-Real Neural Field (2026 | Interspeech 2026 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: Interspeech 2026 author report TR2026-135, 29 September 2026. Access: Full author PDF method, input assumptions and explicit timing/scan future work inspected.

Research contribution. S2RNF learns to convert HRTFs simulated from a person's head mesh into responses closer to measured data. The reviewed paper infers listener-specific neural-field parameters from simulated inputs. It explicitly defers individualized interaural timing and identifies evaluation with photogrammetry-derived meshes as future work; the experiments use high-fidelity meshes available in existing datasets. The contribution reduces one kind of measurement burden under those conditions. [104]

Spatial9 assessment. We see S2RNF as an interesting way to exchange one acquisition burden for another. A geometric scan can be a useful alternative to direct acoustic measurement, but scan quality becomes part of the audio system. Ear detail, incomplete surfaces and registration have consequences that a user may not recognize from a visually acceptable model. We therefore regard capture and timing as integral parts of scan-based personalization rather than incidental setup.

Our verdict. S2RNF fits best in a workflow with dependable geometric capture and an explicit treatment of interaural timing. We value its attempt to connect simulation to measured behavior, while keeping the input quality of the reported work visible. We regard S2RNF as a meaningful contribution to geometry-based personalization, with capture suitability determining adoption: a useful acoustic model cannot make an inadequate scan equivalent to the high-quality mesh it expects.


### A26: SuDaField

SuDaField: Subject- and Dataset-Aware Neural Field for HRTF Modeling (2025 | IEEE Open Journal of Signal Processing 6 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: IEEE OJSP 2025 author manuscript in MERL TR2026-009, 7 January 2026. Access: Full author PDF conversion formulation, unseen-dataset adaptation and conclusion inspected.

Research contribution. SuDaField separates subject-specific and dataset-specific components in an HRTF neural representation. It can alter the dataset component to approximate how a listener's response would appear under another measurement setup. The paper reports promising conversion results and also identifies difficulty adapting its pretrained model to an unseen dataset. Its journal publication is in 2025, despite the associated MERL report carrying a 2026 identifier. [105]

Spatial9 assessment. We consider SuDaField important because personalization data contain measurement conventions as well as anatomy. Treating every difference between datasets as a difference between listeners can give a model the wrong thing to learn. Separating those influences is therefore valuable to us even before it becomes a user-facing feature. We also distinguish harmonizing a collection from improving one person's localization; they occupy different places in the production chain.

Our verdict. SuDaField's strongest architectural role is in the preparation and management of HRTF data. We favor retaining measurement provenance when profiles are combined or converted, rather than presenting a homogenized collection as if it came from one interchangeable protocol. We favor the separation of subject and dataset factors, with the reported unseen-dataset difficulty reinforcing the value of explicit provenance instead of a universal conversion assumption.


### A27: HRTF DDPM

Towards HRTF Personalization using Denoising Diffusion Models (2025 | IEEE ICASSP 2025 peer-reviewed). Status: Peer-reviewed publication, with reviewed author version identified. Reviewed: arXiv:2501.02871v1, 6 January 2025; ICASSP 2025. Access: Full author text conditioning, HUTUBS subject-held-out protocol and evaluation inspected.

Research contribution. The HRTF diffusion study conditions a denoising model on anthropometric information to generate personalized impulse responses. Its reviewed evaluation uses HUTUBS and a leave-one-subject-out procedure, with objective comparison to reference responses. The contribution is the feasibility of generative modeling for listener filters. It concerns the transfer function applied to audio, rather than generation of replacement musical content. [106]

Spatial9 assessment. We regard this paper as a useful reminder that generative processing can act on the listener filter without inventing replacement musical content. That distinction matters to our source-fidelity position. A generated filter still changes the listening experience, and several plausible filters do not automatically identify the best one for a particular person. We therefore treat profile choice and reproducibility as part of personalization, not merely a random-seed detail.

Our verdict. The HRTF diffusion approach belongs naturally in profile acquisition and candidate generation. We favor retaining an approved profile as a stable asset so the same creative scene has a consistent rendering basis across sessions. We regard diffusion as a credible modeling choice for listener filters, with its architectural value tied to selecting and managing a suitable response rather than repeatedly generating a different one during ordinary playback.


## Loudspeaker sound-field control and personal sound zones

We regard loudspeaker control as the physical continuation of spatial authoring. A scene expresses intent, but the room, array and listener positions determine what the installation can realize. The papers in this chapter matter because they address that transition directly. Isolation, soundfield accuracy and perceived spatial placement are different achievements, and we keep those distinctions visible when judging the value of a controller.

Our position favors a clear interface between content and specialized reproduction. A reusable authored scene can coexist with a venue-specific controller, but format compatibility alone does not supply calibration or acoustic control. We see these methods as relevant to installation and reproduction partnerships, with adoption suitability determined by supported movement, array configuration, bandwidth and maintenance effort. That is the architecture we advocate across the wider spatial-audio market.

[Image: Personal sound zones: bright-zone target, dark-zone leakage and a soft gain penalty computed over calibrated transfer paths (code example 14).]


### NAC02: SANN-PSZ

SANN-PSZ: Spatially Adaptive Neural Network for Head-Tracked Personal Sound Zones (2025 | IEEE Transactions on Audio, Speech and Language Processing 33, 2735–2748). Status: Peer-reviewed journal. Reviewed: Published journal PDF; current version 2 July 2025. Access: Full published paper method, experimental band/motion assumptions and discussion inspected.

Research contribution. SANN-PSZ maps listeners' head positions to loudspeaker filters for personal sound zones. Its training uses simulated or combined simulated/measured acoustic transfer functions and studies robustness through augmentation. The reviewed journal experiment uses an eight-loudspeaker linear array over 100–1500 Hz, with horizontal translational movement rather than full head rotation or vertical motion. These conditions are explicit parts of the demonstrated scope. [107]

Spatial9 assessment. We see SANN-PSZ as work on a different responsibility from immersive content authoring: controlling where acoustic energy goes after it leaves the loudspeakers. That distinction is central to our view of the wider market. A well-authored scene still depends on array geometry, listener tracking and the room. We value the position-conditioned controller, while reading its movement and bandwidth limits as part of the supported installation contract.

Our verdict. SANN-PSZ is a relevant specialist reference for head-tracked sound zones under controlled physical conditions. We favor keeping content intent separate from the controller that realizes it, with an explicit connection between the two. That leaves room for specialized reproduction expertise without claiming that a delivery format solves acoustic isolation. We regard SANN-PSZ as a strong integration reference for sound-zone systems, with applicability defined by the array, movement and frequency conditions it supports.


### Code example 14: A sound zone objective must see the calibrated acoustic paths

Illustrative integration pattern. Related reviews: NAC03, NAC05. We judge a loudspeaker controller through the acoustic transfer paths connecting each speaker to each control point. This original PyTorch fragment expresses SANN-PSZ's bright-zone, dark-zone and soft gain-limit terms from equations 5–7. [107, 148, 149, 150]

```python
import torch
import torch.nn.functional as F
def zone_loss(Hb, Hd, g, target_mag, alpha, beta, gmax):
    # Complex H: [frequency, point, speaker]; g: [frequency, speaker].
    # target_mag: [frequency, bright point], real and nonnegative.
    pb = torch.einsum("fml,fl->fm", Hb, g)
    pd = torch.einsum("fml,fl->fm", Hd, g)
    bright = (pb.abs() - target_mag).square().mean()
    dark = pd.abs().square().mean()
    excess = F.relu(g.abs() - gmax).square().mean()
    loss = alpha * bright + (1 - alpha) * dark + beta * excess
    return loss, dict(bright=bright, dark=dark, excess=excess)
```

Reading the example. Hb and Hd contain calibrated or simulated complex transfer functions for disjoint, nonempty zones; g contains predicted speaker gains at the same frequencies. With matching complex dtype/device, the pressure calculation remains differentiable. Alpha balances target magnitude and leakage; beta penalizes gains exceeding gmax. Choose alpha in [0,1], beta nonnegative and a positive calibrated gmax.

Scope. The paper's fourth compactness term, overlap handling, training network and FIR construction are omitted. A soft gain penalty is not a hard actuator limit. Simulated loss does not establish measured isolation, perceptual quality or behavior after listeners or room conditions change.


### NAC03: Flexible bright-zone neural PSZ

Neural personal sound zones with flexible bright zone control (2025 | arXiv author preprint). Status: Preprint; primary conference acceptance not verified. Reviewed: arXiv:2512.10375v1, 11 December 2025. Access: Full author preprint network formulation and simulated-room experimental setup inspected.

Research contribution. Neural personal sound zones with flexible bright zone control uses a three-dimensional convolutional network to derive loudspeaker prefilters from a desired virtual acoustic target. The reviewed preprint varies control grids and reproduction targets. Its experiments use simulated transfer functions in a rectangular room, a circular loudspeaker array and a limited frequency band. A primary conference-acceptance record was not verified, so this review retains its preprint status. [108]

Spatial9 assessment. We like the flexibility to describe a desired acoustic region after training because a listening area is a meaningful control for an installation designer. That moves the interface closer to an intended physical result. The important distinction is between flexibility in specifying a target and reliability across different rooms or arrays. We regard the former as an architectural contribution without treating it as evidence that the latter comes automatically.

Our verdict. This work is useful to our thinking about the boundary between scene authoring and loudspeaker control. A target field can express what the installation wants, while the controller handles how the available hardware attempts to produce it. We favor flexible target specification while reading deployment breadth narrowly: the reported simulated setting supports a controller concept, with room-specific realization remaining an essential part of any adoption decision.


### NAC04: Irregular-array CNN synthesis

Synthesis of soundfields through irregular loudspeaker arrays based on convolutional neural networks (2024 | EURASIP Journal on Audio, Speech, and Music Processing 2024, article 17). Status: Peer-reviewed journal. Reviewed: Published open-access article dated 28 March 2024. Access: Full published article method, simulated and measured-response experiments, and conclusion inspected.

Research contribution. The irregular-array soundfield-synthesis study uses a complex-valued convolutional network to compensate model-based loudspeaker driving signals. It evaluates simulated layouts and a measured-response setup, including arrays made irregular by removing loudspeakers. The model is retrained for each array configuration. The main formulation focuses on two-dimensional reproduction, and the measured-data evaluation uses a restricted acoustic environment and frequency range. [109]

Spatial9 assessment. We value the paper's attention to irregular arrays because installation constraints do not disappear when content has an ideal spatial representation. Neural compensation is relevant precisely at the point where intended sound meets imperfect hardware. The retraining requirement also changes the economics of adoption. A stable venue can amortize configuration work in a way that a frequently reconfigured system cannot, so the same algorithm can have very different practical fits.

Our verdict. We place this method in installation-specific reproduction rather than general scene authoring. Its learned compensation is most attractive when the array and operating region are sufficiently stable to justify calibration and retraining. That favors a clean separation between reusable content and a venue-dependent playback controller. We see a practical role for this method in compensating imperfect speaker geometry, with configuration stability central to its product value and no automatic extension to arbitrary three-dimensional installations.


### NAC05: BSANN

Stereo Audio Rendering for Personal Sound Zones Using a Binaural Spatially Adaptive Neural Network (BSANN) (2026 | IEEE Transactions on Audio, Speech and Language Processing, Early Access). Status: Peer-reviewed journal early-access record from primary author institution. Reviewed: Author manuscript linked from 2026 early-access publication record; reviewed 25 October 2026. Access: Full author manuscript physical modeling, measured-room validation and explicit limitations inspected.

Research contribution. BSANN extends neural personal sound zones to separate stereo programs through ear-optimized loudspeaker filters conditioned on listener positions, physical acoustic modeling and active crosstalk cancellation. Its author manuscript reports measurements in two real rooms and additional static listener arrangements. The paper explicitly separates physical isolation metrics from perceived spatial accuracy and leaves continuous trajectory evaluation and formal subjective testing for future work. [110]

Spatial9 assessment. We consider BSANN particularly relevant because it brings the listener's ears into personal-zone control. Reducing leakage between programs and controlling crosstalk can make shared-space listening more practical. We also preserve the paper's distinction between physical isolation and perceived spatial placement. A system can reduce interference successfully without reproducing every direction intended by the author, and those are separate product promises rather than competing definitions of success.

Our verdict. BSANN is an interesting specialist reproduction layer for delivering separate stereo programs in a shared space. We favor an architecture that lets such a controller handle room- and listener-dependent delivery while the authoring system retains responsibility for source and scene intent. We favor the ear-oriented control strategy and measured-room grounding, with the reported static configurations defining its demonstrated role rather than a complete solution for freely moving immersive listeners.


## Representations and spatial evidence

We distinguish observed acoustic evidence from a plausible inference about missing information. An embedding may make a direction accessible, a language system may answer a grounded question and a codec may preserve channel relationships. These are different contributions. Our assessment gives each representation credit for its demonstrated role and keeps downstream assumptions visible. [111, 112, 113, 114]

Spatial9's position is that scene intelligence depends on the information retained between stages. Directional events, diffuse energy, source identity and timing can matter for different reasons. Compression efficiency, language fluency and a valid scene description are useful attributes, but none independently establishes the audible outcome. We value representations that maintain a clear connection between input evidence, an actionable decision and the experience delivered to the listener.

For how a delivered immersive bitstream is actually structured, see the IAMF technical deep dive.


### R10: SARL

Probing Spatial Structure in Pretrained Audio Representations (2026 | Interspeech 2026). Status: Acceptance stated by authors and dataset record. Reviewed: arXiv 2606.05544v2. Access: Full text: introduction and controlled probing methodology.

Research contribution. SARL probes pretrained audio representations for source and room attributes under controlled conditions. The work separates properties such as source class, direction and distance from room-related information to examine what frozen representations make accessible. The reviewed result concerns representation probing; it is not an end-to-end demonstration of room reconstruction or professional mix planning. [111]

Spatial9 assessment. We regard spatial representation as a constraint on everything a planner can reliably decide. Fluent language does not create acoustic evidence that the representation has discarded. Probing is useful because it exposes which source and room attributes are accessible, while accessibility itself remains different from reliable downstream use. This distinction is fundamental to our interpretation of audio-language systems.

Our verdict. SARL is a strong diagnostic reference for selecting and understanding perception layers. Its contribution is especially relevant when a system makes confident spatial statements from generic audio features. For Spatial9, we place acoustic grounding ahead of the polish of an explanation. We view representation analysis, planning behavior and final audible quality as separate responsibilities that have to agree for a scene decision to be trustworthy.


### R11: Spatial audio LLM

Can Large Language Models Understand Spatial Audio? (2024 | Interspeech 2024). Status: Conference proceedings. Reviewed: https://www.isca-archive.org/interspeech_2024/tang24d_interspeech.html. Access: Full text: methods and experimental task definitions.

Research contribution. Can Large Language Models Understand Spatial Audio? combines speech representations with explicit spatial information from first-order Ambisonics. The Interspeech paper studies localization, far-field recognition and direction-informed extraction using controlled speech data. Its design is relevant because spatial cues are supplied alongside semantic audio features. The experiments establish progress on those tasks, rather than on musical arrangement or a producer's aesthetic judgment. [112]

Spatial9 assessment. We see explicit spatial features as a sound architectural response to the limits of semantic audio representations. Language provides an interface and a reasoning resource, while acoustic cues carry the positional evidence. The paper makes it unreasonable to dismiss all audio-language systems as spatially incapable. It also gives no basis for equating success on speech tasks with musical judgment.

Our verdict. This work is an important bridge between spatial signal processing and language-based interaction. Spatial9's position is to preserve the connection between a spoken or written instruction and the acoustic facts that justify its execution. We distinguish understanding a direction, extracting a source and choosing an expressive arrangement. Their relationship matters to a product, but evidence for one capability does not automatically establish the others.


### R12: FOA Tokenizer

FOA Tokenizer: Low-bitrate Neural Codec for First Order Ambisonics with Spatial Consistency Loss (2025 | arXiv; submitted to ICASSP 2026). Status: Preprint; acceptance not independently established. Reviewed: arXiv 2510.22241v1. Access: Full text: architecture and spatial consistency loss.

Research contribution. FOA Tokenizer develops a low-bitrate neural representation for four-channel, 24 kHz first-order Ambisonics, with a spatial-consistency loss. The reviewed loss masks low-energy and highly diffuse regions when estimating directional consistency. Its manuscript is a preprint; the stated submission to ICASSP does not by itself establish acceptance. The task is compression and representation, not object editing or production-format delivery. [113]

Spatial9 assessment. We view spatially informed tokenization as an enabling technology, with representation efficiency and delivery interoperability treated separately. The loss's treatment of low-energy and diffuse regions is particularly relevant to immersive production because ambience can carry much of the sense of space. It identifies an evidential boundary; it does not establish that the codec audibly damages those regions.

Our verdict. FOA Tokenizer is a relevant internal-representation reference with a defined first-order, bandwidth and training scope. Spatial9's position gives diffuse energy, reverberant tails and scene continuity explicit status alongside directional events. An efficient token stream is valuable when those intended attributes survive its use. It remains distinct from an interchange format or an editable object scene, and our assessment keeps those engineering roles separate.


### Code example 15: Penalize directional damage in a spatial codec

Illustrative integration pattern. Related reviews: R13, R15. FOA Tokenizer adds a direction-sensitive objective to its codec loss. This original transcription of the paper's weighted cosine penalty exposes the essential choice: energetic, less diffuse reference regions contribute more strongly to the directional constraint. [113]

```python
import torch
def directional_loss(i_ref, i_hat, energy, diffuseness,
                     min_energy=1e-6, max_diffuseness=0.95):
    # i_*: [time, frequency, 3]; other inputs: [time, frequency].
    i_ref, i_hat = i_ref.float(), i_hat.float()
    energy, diffuseness = energy.float(), diffuseness.float()
    dot = (i_ref * i_hat).sum(dim=-1)
    norm = i_ref.norm(dim=-1) * i_hat.norm(dim=-1)
    cosine = (dot / (norm + 1e-8)).clamp(-1.0, 1.0)
    mask = (energy > min_energy) & (diffuseness < max_diffuseness)
    weight = mask * energy * (1.0 - diffuseness)
    return (weight * (1.0 - cosine)).mean()
```

Reading the example. The intensity vectors are real, with matching normalization and sign; energy and diffuseness come from the reference field. Float32 conversion keeps epsilon representable for half-precision inputs. The mean covers all time-frequency bins, matching the paper. Its stated thresholds are shown; epsilon is an illustrative numerical stabilizer.

Scope. This is one loss term, not the codec or an official implementation. It assumes valid, aligned tensors and diffuseness in [0, 1]. Silence and diffuse regions provide little directional supervision, so reconstruction and perceptual objectives remain necessary.


### R13: SpatialCodec

SpatialCodec: Neural Spatial Speech Coding (2024 | IEEE ICASSP 2024; preprint 2023). Status: Conference acceptance stated in author manuscript. Reviewed: arXiv 2309.07432v2, 9 July 2024. Access: Full text: framework, data construction and evaluation.

Research contribution. SpatialCodec separates coding of a reference microphone channel from coding of relative spatial information. Its full manuscript evaluates synthesized reverberant array recordings with a single speech source and uses spatial similarity and beamformed quality measures. This task preserves useful array information under compression; it is distinct from encoding an artist's object-based master or supplying an editable immersive session. [114]

Spatial9 assessment. We value the explicit coding of relationships between channels. Spatial meaning depends on those relationships, so acceptable quality in each isolated channel does not settle the quality of the combined scene. The single-source speech setting gives the work a clear scope. Dense music, independent object control and other array geometries introduce additional requirements rather than retroactively invalidating that contribution.

Our verdict. SpatialCodec is a relevant spatial-representation and compression reference, with its strongest evidence in the published array-speech task. For Spatial9, the architectural lesson is to preserve interchannel relationships as first-class information. We do not classify the model as a music upmixer or an object-mastering format. Its value lies in how it treats spatial information during coding and what that suggests about the interfaces between processing stages.


### NROOT08: L3DAS22

L3DAS22 Challenge: Learning 3D Audio Sources in a Real Office Environment (2022 | IEEE ICASSP 2022). Status: Conference publication; DOI 10.1109/ICASSP43922.2022.9746872. Reviewed: arXiv 2202.10372v1. Access: Full text: measured impulse responses, dataset synthesis and task definitions.

Research contribution. L3DAS22 supports three-dimensional speech enhancement and sound-event localization/detection using Ambisonic data. Its scenes are synthesized with impulse responses measured in a real office, together with clean source recordings. This matters methodologically: measured room acoustics and fully natural scene recordings are different forms of realism. The challenge is designed around office-like tasks, rather than immersive music mastering. [115]

Spatial9 assessment. We regard precise descriptions of dataset realism as essential to interpreting capability. Measured propagation, synthesized mixtures and naturally occurring recordings provide different kinds of evidence. L3DAS22 makes a useful contribution by combining controlled sources with real measured acoustics. Its office tasks are relevant to perception and enhancement, while immersive music brings additional assumptions about material, overlap and creative intention.

Our verdict. L3DAS22 is a valuable reference for grounded three-dimensional perception and enhancement research. Spatial9's assessment keeps its source construction and task scope explicit. We distinguish what a model learns about acoustics from what it learns about the arrangement of events. Controlled data supports diagnosis, and natural production material adds different challenges; neither category alone carries a universal claim of real-world readiness.


## Evaluation and the meaning of a quality claim

We regard evaluation as a set of instruments with defined purposes. A reference metric diagnoses damage, a distributional distance compares collections, and an audiovisual benchmark examines alignment. Creative preference introduces another question. Our position is to assign each result the meaning its assumptions support and resist turning an available number into a universal judgment. [116, 121, 122]

For Spatial9, content fidelity, spatial behavior, perceptual experience and revision usefulness remain visible dimensions of quality. Disagreement between them can reveal a real tradeoff. The surveys and loss review help organize those distinctions, while individual metrics provide focused measurements. We see a clear account of what changed and why it matters as more useful than a single unexplained ranking. [123, 124]

[Image: Why one quality score is not enough: spatial quality is a profile of distinct attributes, each measured by instruments with their own validation domains.]


### R03: SAQAM

SAQAM: Spatial Audio Quality Assessment Metric (2022 | Interspeech 2022). Status: Conference proceedings. Reviewed: https://www.isca-archive.org/interspeech_2022/manocha22b_interspeech.pdf. Access: Full text: methodology and training setup.

Research contribution. SAQAM learns distances between binaural signals along listening-quality and spatialization-quality dimensions. Its training setup uses speech with simulated degradations. Separating those dimensions is useful because a signal can retain intelligibility while losing spatial character, or gain apparent width while acquiring artifacts. The reviewed methodology supports a complementary assessment instrument; it does not validate every possible music-production use. [116]

Spatial9 assessment. We value SAQAM's separation of listening quality from spatialization quality. This matches our view that an attractive spatial impression can coexist with source damage, and that clean audio can still have an unsuitable image. Its speech-centered evidence limits the weight we assign it in music-production decisions. Dense instrumentation and deliberate decorrelation introduce different interpretive questions; that boundary does not establish a defect in the metric.

Our verdict. SAQAM belongs in a diagnostic quality profile, with its validation domain visible. We do not treat its scalar outputs as authority over a producer's creative choice. For Spatial9, the practical value is distinguishing kinds of change and directing attention to possible problems. Source identity, musical intent and listener preference remain separate dimensions of our assessment because they answer questions the metric was not designed to settle.


### R04: AMBIQUAL 2018

AMBIQUAL – a full reference objective quality metric for ambisonic spatial audio (2018 | QoMEX 2018). Status: Conference publication. Reviewed: https://research.google/pubs/ambiqual-a-full-reference-objective-quality-metric-for-ambisonic-spatial-audio/. Access: Author institution abstract and metadata.

Research contribution. The 2018 AMBIQUAL paper extends reference-based audio-quality assessment to Ambisonic B-format signals. The author-institution description presents listening quality and spatial localization as related assessment targets. This review accessed the primary abstract and bibliographic record, rather than completing a detailed manuscript read. Accordingly, the assessment here is confined to the documented task and its implications, without attributing unverified experimental weaknesses. [117]

Spatial9 assessment. We regard AMBIQUAL as important prior art for evaluating spatial localization alongside signal fidelity. The defining assumption is the existence of a meaningful reference. That is a useful basis for assessing damage introduced by processing or delivery. A creative rearrangement has a different relationship to its source: some changes are intentional, and fidelity to an unknown spatial original cannot be assumed.

Our verdict. AMBIQUAL is relevant to reference-based quality assurance, with a bounded role in creative production. For Spatial9, the governing question is which properties the operation is supposed to preserve. We accept reference comparison as evidence about those properties, while keeping artistic interpretation explicit. This assessment concerns the documented purpose of the method; it assigns no unverified numerical advantage or implementation-specific weakness to the 2018 work.


### R05: AMBIQUAL 2020

AMBIQUAL: Towards a Quality Metric for Headphone Rendered Compressed Ambisonic Spatial Audio (2020 | Applied Sciences 10(9), 3188). Status: Journal article; extension of AMBIQUAL. Reviewed: DOI 10.3390/app10093188. Access: Author institution abstract and metadata.

Research contribution. The 2020 AMBIQUAL journal article extends the earlier work to headphone-rendered compressed Ambisonics, including first- and third-order material and fixed or moving sources. Its institutional abstract reports comparison with subjective judgments under Opus compression. It is a journal extension of the same research family, not an unrelated model. Full-text retrieval was blocked during this review, so detailed experimental claims remain outside this entry. [118]

Spatial9 assessment. We see the journal extension's focus on compressed, headphone-rendered Ambisonics as directly relevant to delivery. The listener receives the decoded and rendered signal, so quality at an earlier stage cannot stand in for the final experience. Our assessment also keeps the impairment family explicit: coding damage, learned reconstruction and an intentional mix change have different meanings even when they affect similar acoustic cues.

Our verdict. This AMBIQUAL work strengthens the case for treating encoding and rendering as part of the quality boundary. In Spatial9's view, a successful spatial production retains its intended behavior through the actual delivery path. Reference-based measurements are useful within the codec and rendering conditions they describe. We reserve conclusions about other formats, object semantics and creative choices for evidence that addresses those questions directly.


### R06: QASTAnet

QASTAnet: A DNN-based Quality Metric for Spatial Audio (2026 | IEEE ICASSP 2026; preprint first posted 2025). Status: Conference DOI verified in arXiv metadata. Reviewed: arXiv 2509.16715v2. Access: Full text: methods, dataset, evaluation, conclusion.

Research contribution. QASTAnet combines auditory features with a compact neural predictor of subjective spatial-audio quality. The reviewed version concerns third-order Ambisonics codec impairments. Its conclusion explicitly identifies other codecs, order truncation, packet loss and additional channel layouts as future extensions. Those boundaries are useful because they identify where its reported evidence is concentrated rather than implying a general-purpose judge of all immersive production. [119]

Spatial9 assessment. We prefer a focused quality metric with a clear domain to a broad score whose meaning is unclear. QASTAnet's concentration on a defined coding problem makes its evidence interpretable. For Spatial9, coding artifacts, source damage, spatial displacement and dissatisfaction with a creative decision are distinct categories. A good prediction for one category does not resolve the others, even when a product combines all of them in one pipeline.

Our verdict. QASTAnet is a useful specialist reference for Ambisonic codec-quality assessment. We place it within a larger quality profile rather than giving it final authority over an immersive mix. Its production relevance is strongest where the representation and impairment conditions match the published task. Beyond that domain, our position is to retain the question as open rather than convert an available score into an unsupported verdict.


### R07: SPAUQ

Quantifying Spatial Audio Quality Impairment (2024 | IEEE ICASSP 2024; preprint 2023). Status: Conference proceedings. Reviewed: arXiv 2306.08053v3. Access: Full text: proposed method and limitation discussion.

Research contribution. Quantifying Spatial Audio Quality Impairment proposes a decomposition that distinguishes spatial errors, including gain leakage and relative delay, from other signal errors. The paper acknowledges that frequency-dependent room and pinna effects cannot be completely disentangled from equalization distortion within this formulation. It is a diagnostic framework for impairment, rather than a predictor of which artistic spatial arrangement listeners will prefer. [120]

Spatial9 assessment. We consider causal diagnosis more useful to engineering than an unexplained aggregate score. Spatial leakage, timing changes and spectral coloration can enter at different stages and require different remedies. The decomposition in this paper supports that way of thinking while retaining an important limit: room and pinna effects can interact with what the formulation classifies as equalization. A diagnostic remains useful when its ambiguity is understood.

Our verdict. Spatial impairment decomposition is a valuable engineering lens for locating damage within a production chain. For Spatial9, the central implication is to preserve visibility into intermediate representations and interpret each error category in context. We do not read a lower impairment score as a judgment of musical taste. The method's strongest role is explaining a change in the signal, with listening and task-specific interpretation supplying its practical meaning.


### R08: FAD adaptation

Adapting Frechet Audio Distance for Generative Music Evaluation (2024 | IEEE ICASSP 2024). Status: Conference publication. Reviewed: arXiv 2311.01616v2, 5 March 2024. Access: Full text: methodology, embedding choices and evaluation.

Research contribution. The FAD adaptation study examines how sample size, embedding choice and reference-set selection affect generative music evaluation. It proposes ways to address sample-size bias and investigates perceptually relevant configurations. The full text makes clear that a distributional score depends on how that distribution is represented. This is adjacent evaluation evidence, not a dedicated spatial localization metric. [121]

Spatial9 assessment. We distinguish the plausibility of a music distribution from the faithful treatment of a particular recording. An output can resemble high-quality music while altering a detail its creator regards as essential. Equally, an unusual artistic choice can fall outside a reference collection without being poor work. FAD's sensitivity to representation and reference selection makes these distinctions central to interpreting the number.

Our verdict. FAD is relevant background for generative quality assessment and has a limited role in judging a source-preserving spatial transformation. Spatial9's position gives priority to the identity, timing and intended treatment of the supplied work. Distributional resemblance is an additional perspective, not a replacement for that obligation. We regard transparent embedding and reference choices as necessary to understanding any FAD-based argument.


### R09: SAVGBench

SAVGBench: Benchmarking Spatially Aligned Audio-Video Generation (2026 | IEEE ICASSP 2026; first preprint 2024). Status: Conference acceptance and author repository verified. Reviewed: arXiv 2412.13462v2. Access: Full text: task, dataset and metric construction.

Research contribution. SAVGBench evaluates spatial alignment between generated audio and video using audiovisual material and detector/localizer-based measurements. Its methods connect visible source evidence to estimated sound direction. That is an important evaluation target for audiovisual generation, but the measurement also depends on the capabilities and assumptions of the detection and localization systems used to construct it. [122]

Spatial9 assessment. We value spatial audiovisual alignment as a clearly defined capability. We also distinguish physical correspondence from narrative intention: an off-screen voice, displaced effect or nondiegetic score can be correct for the brief without matching a visible object. Our concern is therefore how a benchmark's assumptions are applied to a production claim. Detector and localizer behavior also belongs inside the interpretation of an automated score.

Our verdict. SAVGBench is a useful reference for spatially aligned generation, with a specific editorial scope. Spatial9's position is to make the intended relationship between image and sound explicit before assigning quality. Visible-source correspondence, off-screen placement and artistic displacement are different instructions. A production system earns trust by respecting the chosen relationship, including deliberate exceptions to literal visual alignment.


### R14: ASAudio survey

ASAudio: A Survey of Advanced Spatial Audio Research (2025 | IJCNLP-AACL 2025). Status: Published survey. Reviewed: https://aclanthology.org/2025.ijcnlp-long.25.pdf. Access: Official abstract, taxonomy and bibliographic discovery.

Research contribution. ASAudio surveys spatial-audio generation, understanding, representations, datasets and evaluation using a task-oriented taxonomy. The official publication is useful for mapping research branches and tracing references across communities. As a 2025 survey, it necessarily precedes some of the 2026 work included here. A taxonomy organizes evidence; it does not provide a controlled ranking of every method it describes. [123]

Spatial9 assessment. We regard a task-oriented map as the foundation of useful criticism. Binaural synthesis, source separation, acoustic reconstruction and scene reasoning can contribute to the same product while solving different problems. ASAudio's breadth supports that distinction. Our own position extends the map into production decisions: what the creator can control, what the system infers and what the delivery representation retains.

Our verdict. ASAudio is a valuable organizing reference for the field. A survey's authority comes from its synthesis and traceable scope, not from serving as a performance leaderboard. Spatial9's assessment uses that same discipline to distinguish competing approaches from complementary components. The resulting research map gives each contribution a place and makes criticism more useful than a list of nominal competitors.


### R15: Spatial loss review

Loss functions incorporating auditory spatial perception in deep learning -- a review (2025 | IEEE I3DA 2025). Status: Conference DOI linked in author manuscript. Reviewed: arXiv 2506.19404. Access: Primary manuscript: stated scope and review findings.

Research contribution. The spatial-loss review examines perceptually motivated objectives for binaural reproduction. It identifies a strong emphasis on localization cues and less developed treatment of room-related attributes. Its stated scope excludes room-impulse-response losses and general spectral-temporal characteristics. Those exclusions define a focused review; they are not evidence that the omitted dimensions are unimportant or absent from all research. [124]

Spatial9 assessment. We interpret the review as support for a multidimensional account of spatial quality. Intended direction, externalization, reverberation and timbral integrity can diverge. An objective emphasizing one cue family cannot carry the meaning of the entire listening experience. The review's exclusions are useful because they make that boundary visible instead of disguising a partial perspective as a complete perceptual model.

Our verdict. This loss-function review is an important guide to what optimization emphasizes and what it can leave unresolved. For Spatial9, our quality philosophy keeps the components of the experience visible and connects them to the creative brief. A deliberately unnatural effect may be appropriate, while a physically plausible scene may be musically unsuitable. We therefore judge an objective by the attribute it addresses and the consequences of prioritizing it.


## Why one quality score is not enough

The evaluation chapter above leads to a simple rule. Fidelity, localization, externalization, immersion and artistic preference are related but distinct. SAQAM, AMBIQUAL, QASTAnet, spatial impairment decomposition and perceptual-loss research provide instruments with different domains. Our assessment favors an interpretable profile of outcomes over a single unexplained score, especially when improvements in one attribute coincide with a tradeoff in another. [116, 117, 118, 119, 120, 124]


## What the papers collectively change

We see spatial intelligence as a set of distinct capabilities. Inferring an existing source's location, generating a requested event, describing a scene, modeling propagation and choosing intentional movement answer different questions. Our assessment gives each capability a clear role. This task map is central to understanding where approaches compete, where they complement one another and where a product has additional responsibilities.

BAT, the DCASE question-answering work and Spatial-Omni broaden the field beyond rendering and upmixing. We regard acoustic grounding as the decisive issue when language becomes part of a spatial workflow. A useful response connects the user's request to evidence in the scene. Spatial understanding, reasoning over that understanding and executing an edit remain distinct capabilities in our assessment. [31, 33, 34]

Generation and preservation impose different obligations. A task that creates missing material can benefit from plausible inference, while a task that transforms an approved recording has a stronger commitment to its identity. Our position makes that distinction explicit at the operation level. Source retention, reconstruction, generation and mixing into a sound field each have consequences for what remains audible and editable.

The catalog already includes motion-aware generation, streamable binaural synthesis and acoustic rendering under spatial transformations. We credit those advances directly. Our production perspective focuses on continuity between capabilities: a source is identified, assigned a role, moved, revised, encoded and played back. The value of an individual stage depends partly on whether its information remains useful to the stages that follow. [45, 53, 88]

HRTF personalization, room reconstruction and sound-zone control describe different relationships between signal, listener and environment. We treat calibration, geometry, physical constraints and changed conditions as part of each method's operating contract. Our assessment distinguishes a general spatial effect from the reconstruction or control of a particular listening situation. That distinction defines where each approach has practical value. [98, 99, 88, 107]

Our final synthesis concerns evidence. A focused metric, a controlled acoustic experiment and a qualitative workflow study supply different kinds of knowledge. We value all three while retaining their scope. Spatial9's position connects these perspectives through the creator's task: what the method changes, what remains controllable and what the listener receives. The criteria below make that judgment explicit. [116, 119, 9, 10]


## What this means for Spatial9

Spatial9's documented architecture connects analysis, placement decisions, a scene engine and rendering or export. The public workflow material also describes motion controls and multiple delivery routes. We use that structure to locate each research contribution within a production system. The central architectural question is how the information and decisions at one stage remain meaningful at the next. [129, 130, 133]

Our source-fidelity position concerns the identity and intended treatment of supplied material. Separation, resampling, equalization, spatial filtering and coding can all change a signal, so preservation needs a precise meaning. Bit identity, perceptual transparency and recognizable source identity are different properties. We regard an explicit account of the transformation as essential to understanding the result. [129]

The documented conversion route includes third-order Ambisonics, while the Acoustic Digital Twin demonstration retains its research-preview status. We distinguish an internal object scene, a rendered sound field and an exported object representation by the information each carries. Acoustic reconstruction has a further physical obligation to the room being represented. Those distinctions shape our verdicts on integration, delivery and research priorities. [132, 131]

Our architecture position is to retain clear responsibilities for perception, decisions, scene state, rendering and handoff. A useful scene state identifies the source, the scope of an operation and the intention behind it. A useful handoff makes the delivered representation unambiguous. These are the principles through which we interpret the literature, with product descriptions and research findings kept attributable to their respective sources.

[Image: The seven principles we apply to every paper in this review and to our own architecture.]

| Area | Our position | Practical implication |
| Placement and planning | Instructions need acoustic grounding | The source, decision and affected interval remain identifiable |
| Source fidelity | Preservation has an explicit meaning | Identity, transparency and bit identity are distinguished |
| Scene revision | A local request has a local scope | Unrelated choices remain part of the creator's retained intent |
| Musical motion | Movement serves the composition | Stationary choices, timing and path design have separate roles |
| Export behavior | The delivered representation defines the handoff | Object state, sound-field rendering and player support are explicit |
| Room reconstruction | Calibration belongs to the model's operating contract | Acoustic inference is distinguished from a measured room match |
| Listener variation | A spatial result depends on reproduction conditions | HRTF fit, devices and listener behavior affect interpretation |


## The fourteen criteria behind our assessments

Our verdicts use fourteen production criteria. They explain why an approach is compelling, bounded, integration-dependent or unsuitable for a particular claim. A method does not need to satisfy every criterion to make a valuable contribution. The relevant criteria follow from the task it promises to solve and the role it occupies in the complete workflow.

We distinguish component capability from workflow usefulness. A separator, planner, renderer, room model or codec has a defined responsibility. A production system connects those responsibilities with user decisions, correction and delivery. Our assessment can therefore recognize a strong component while identifying the work required around it, without treating those additional responsibilities as failures of the original research.

These criteria are the reasoning framework for the paper's assessments and verdicts. They provide a consistent vocabulary for technical fit, evidence boundaries and creator value. They are not a table of unpublished numerical outcomes. The central requirement is that the meaning of a claim matches the behavior, representation and listening condition it describes.


### 1. Task and input parity

We assess a result in relation to the information the method receives: source assets, metadata, room or listener information and signal conventions. Clean stems and a stereo mixture are different inputs. Exact acoustic calibration and a generic assumption are different resources. A practical baseline and a calibrated reference can both be informative, provided their roles remain clear. Our verdict identifies which advantages come from the method and which depend on its operating information.


### 2. Content fidelity

We define fidelity through the properties the operation is meant to preserve. Missing events, changed identity, altered timing and unintended timbral changes matter differently from an intentional spatial filter. Our assessment distinguishes identity preservation, perceptual transparency and bit identity. It also considers the transitions between separation, scene processing, rendering and coding. A source-

preservation claim is meaningful when its protected properties and permitted transformations are explicit.


### 3. Separation and identity

We judge separation by its relationship to the desired unit of control. A broad stem can be useful while still combining sources the creator wants to move independently. Leakage, missing material, duplication and identity changes affect that relationship. Promptable extraction adds another form of control, with its own ambiguity. Our position gives the source inventory and the stability of those identities the same importance as a familiar separation score.


### 4. Audible direction height and distance

We distinguish the spatial command from the perceived result. Coordinates describe intended geometry, while the rendered cues and listener response determine audible direction, height and distance. An estimator adds its own uncertainty. Our assessment keeps azimuth, elevation, front/back distinction and distance separate because success on one does not settle the others. The scene description and its audible consequence are complementary parts of the production claim.


### 5. Motion and musical intention

We assess motion through its musical purpose, timing, continuity and controllability. A stationary source can be the appropriate choice. An expressive path needs a suitable onset, duration and relationship to the surrounding material. Listener movement introduces another frame of reference. Our position treats event selection, trajectory design and revision as connected but distinct responsibilities, with artistic appropriateness retained alongside geometric smoothness.


### 6. Complete song and scene continuity

We consider coherence across a complete work part of production usefulness. Repeated sections, sparse openings, dense peaks, silence and decaying endings expose different demands on a system. Source identity and room behavior also have to remain interpretable through edits and boundaries between processing segments. A method designed for short clips retains its contribution; our assessment identifies the additional continuity responsibilities of the workflow around it.


### 7. Downmix and alternate playback

We regard alternate playback as part of the delivery relationship. Downmixing can change balance, cancellation, transients and low-frequency behavior, while an intentional remix can appropriately differ from its stereo source. The chosen matrix or renderer determines the interpretation. Our position keeps the intended fallback experience explicit and avoids assuming that a convincing immersive preview establishes equally convincing behavior on every supported arrangement.


### 8. Latency and production cost

We judge speed in relation to the use case. Throughput and real-time factor describe processing capacity; buffering, lookahead and transport affect interaction delay. Offline production also includes import, correction, rendering and export. Our assessment values time saved toward an accepted result, including recovery from an unsuitable suggestion. Hardware, precision and operating conditions belong to the meaning of a performance claim.


### 9. Room modeling

We regard calibration and environmental validity as part of an acoustic model's contract. Geometry, material assumptions, microphone conditions and room-specific data determine what its predictions mean. A useful reconstruction describes a particular acoustic situation, while artistic reverberation serves a different purpose. Changed furniture, openings or listener positions can alter that situation. Our assessment of Spatial9's acoustic research direction retains these physical obligations and the documented preview status. [131]


### 10. HRTF and listener variation

We assess personalization through both the information required from a listener and the listening property it addresses. Filter-reconstruction error, front/back distinction, elevation and externalization are related but different outcomes. Headphone response and alignment also affect interpretation. Our position gives listener variation a visible place in the quality profile and distinguishes generic compensation from an individualized acoustic representation.


### 11. Export and playback verification

We treat the actual encoded representation as the handoff contract. Channel ordering, normalization, timing, tails, loudness and object or scene elements determine what the receiver can use. Rendering objects into a field changes their independent editability even when spatial behavior remains audible. Our assessment distinguishes specification capability from implementation and player support, with both contributing to the usefulness of a delivered file. [125, 126]


### 12. Listening studies

We interpret listening evidence through the question participants answer. Reference-based impairment, preference among creative arrangements and compliance with an artistic brief are distinct tasks. Listener experience, devices, presentation conditions and response variation affect the conclusion. Spatial9's position keeps those dimensions visible and gives a protocol its proper scope. A familiar method name does not turn every rating into a universal statement of spatial quality. [127, 128]


### 13. Metric validity

We assign each objective metric a defined role. Its validation domain and the defect it measures determine how much authority its score carries in a new setting. Disagreement with a relevant listener judgment can reveal a limitation or a different measured attribute. Our assessment treats diagnostic value, perceptual prediction and creative preference separately. A metric can be useful without supplying a complete ranking of immersive systems. [116, 119, 121]


### 14. Editable delivery and recovery

We regard revision and recovery as central to creative control. Moving one source, constraining an interval, undoing a decision and reopening a scene describe the creator's ongoing relationship with the result. The receiving representation determines which of those actions remain possible after delivery. Our position gives preserved intent and localized correction explicit architectural value, extending the assessment beyond the initial output to the next meaningful decision.


### How we interpret comparative evidence

We interpret evidence at the level of the claim. Many adjacent frames from one song do not represent the same variation as independent songs. Repeated listening judgments carry both participant and stimulus effects. Our assessment gives uncertainty, effect size and failure distribution a place beside average outcomes, because those details determine the practical strength of a comparison.

We distinguish a technical failure from an artistic disagreement. A missing source, discontinuous trajectory or corrupted handoff can violate a declared requirement. Two coherent arrangements can satisfy the same brief while appealing to different listeners. Our verdict preserves that distinction. Technical compliance, immersion and creative preference can diverge without making the evidence meaningless.

We retain the identity of the contribution being judged. A faithful reproduction, an adapted implementation and an integrated product supply different evidence. A third-party separator's result remains attributable to that separator even when it is part of a larger workflow. Our interpretation credits improvements at the stage where they arise and distinguishes component evidence from the usefulness of the complete session.

We consider failure boundaries useful technical information. A method can be valuable for a defined range of material, rooms, listeners or devices while leaving other conditions unresolved. Our assessments state those boundaries directly. This produces a more actionable account of suitability than treating every paper as a candidate for universal superiority.


## Positioning Spatial9 across the market

Spatial9's position is an AI spatial production system centered on controllable creative intent. This connects the research branches through the creator's task. Expressive results, dependable revision and a clear handoff give the position practical meaning, while the research literature supplies specialized contributions that inform each part of the workflow.

Our editorial center is the complete creative decision: the available sources, the intended arrangement, the requested revision and the delivered experience. That scope makes the relationship between model capability and production usefulness visible. The technical value of the position lies in explaining what changes, what remains stable and what the recipient receives.

For music, we emphasize source identity, musical restraint and coherent whole-song revision. For film and video, we emphasize narrative relationships, visible and off-screen sources, and camera changes. Interactive environments add listener movement and responsive scene behavior. Acoustic reconstruction and loudspeaker control carry distinct physical obligations involving calibration and environmental validity. These are different applications of the same commitment to explicit intent.

We recognize that a specialized model can be the right tool for a specialized task. Our integration perspective gives advances in separation, perception and rendering a clear place while locating product value in the decisions between those stages. Attribution strengthens that account: the contribution of a component and the contribution of a complete workflow can both be substantial.

Our public voice gives credit to the contribution, states a precise technical judgment and explains its consequence for creators. We criticize assumptions that limit a claim and interfaces that lose useful information. We also apply those standards to Spatial9's own priorities. The result is a position that readers can understand, challenge and use in their own engineering or creative decisions.


## Standards and delivery conventions

Standards are included for delivery and evaluation context, not counted as academic models. For a practical look at the files themselves, see what an ADM BWF file is and how Eclipsa Audio compares with Dolby Atmos.


### D01: Audio Definition Model

ITU-R BS.2076-3 Audio Definition Model (2025, ITU-R Recommendation BS.2076-3, Standard; February 2025 revision).

ADM defines a structured description of audio and associated metadata. Its relevance is the representation of a production's content and rendering information; the existence of the standard does not establish that every exporter or downstream tool preserves every intended relationship. [125]

Our position is that an ADM handoff is defined by the relationship between its audio and metadata. Identifiers, timing, coordinate conventions and source associations determine what the receiving tool can reconstruct and edit. A successful file open is only one aspect of that contract. For Spatial9, the meaningful delivery claim identifies which elements remain independent and how the recipient interprets their intended behavior.


### D02: Immersive Audio Model and Formats

Immersive Audio Model and Formats v2.0.0 (2026, Alliance for Open Media IAMF 2.0.0, Final deliverable dated 24 August 2026).

IAMF 2.0 supports channel-, scene- and object-based audio structures. That specification-level capability must be separated from support in a particular encoder, export profile, player or distribution platform. The reviewed specification is the approved 2.0 version current at the search cutoff. [126]

Our position separates the expressive capacity of IAMF from the structure emitted by a particular workflow. Rendering independent objects into an Ambisonic field is a meaningful representation choice: spatial behavior can remain audible while independent object editability changes. For Spatial9, the delivered element type, parameter timing, loudness behavior and decoder support define the practical contract. The format name alone does not describe those properties.


### D03: Assessment of intermediate audio quality

ITU-R BS.1534-3 Method for the subjective assessment of intermediate quality level of audio systems (2015, ITU-R Recommendation BS.1534-3, Subjective assessment recommendation).

BS.1534 describes a method commonly known as MUSHRA for appropriate reference-based assessments of intermediate audio quality. Its role is to structure an impairment experiment, with a defined reference and comparison conditions, rather than to declare a unique correct artistic mix. [127]

We regard MUSHRA as an appropriate tool when the question concerns impairment relative to a defined reference. Creative arrangements introduce a different question because several outcomes may satisfy the brief. Spatial9's position keeps reference quality, preference and intent compliance distinct. The listening task, comparison conditions and interpretation of ratings determine the strength of a conclusion; a familiar protocol name cannot substitute for that alignment.


### D04: Assessment of small audio impairments

ITU-R BS.1116-3 Methods for the subjective assessment of small impairments in audio systems (2015, ITU-R Recommendation BS.1116-3, Subjective assessment recommendation).

BS.1116 addresses subjective assessment of small audio impairments. It belongs in a research program concerned with subtle processing or delivery differences, where the experimental design and listening conditions need to match that purpose. [128]

We regard BS.1116 as relevant to claims about subtle impairment under suitable listening conditions. Its role is distinct from judging which creative spatial arrangement someone prefers. For Spatial9, transparency claims depend on the material, listeners, devices and sensitivity of the comparison. An undetected difference has a bounded meaning, and our assessment retains that meaning rather than treating it as universal signal identity.


## The Spatial9 conclusion

Our assessment of the literature is that spatial audio now has a substantial set of specialized capabilities across perception, generation, acoustic modeling and control. Their diversity is valuable. The production challenge is connecting those capabilities while retaining the meaning of sources, decisions and representations through ambiguity, revision and delivery.

Spatial9's position centers on that continuity. A creator needs to understand which source is controlled, what an instruction changes, which intentions remain intact and what the recipient receives. We value learned processing where it improves that relationship and explicit structure where it keeps decisions understandable and revisable. Fidelity, spatial behavior, creative preference and practical control all remain part of the judgment.

The 124 individual verdicts support a consistent technical view: useful intelligence makes spatial decisions more manageable across the complete life of a scene. No single research family owns every part of that process. Our commitment is to evaluate each contribution precisely, preserve the distinctions that matter and connect technical progress to the creator's ability to shape an intentional listening experience.

If you want to see how these principles show up in practice, explore how AI agents can drive spatial decisions through the Spatial9 MCP and API.


## Frequently asked questions


### What is AI spatial audio?

AI spatial audio is the use of machine learning to analyze, create, place, render or deliver sound in three-dimensional space. In this review it spans twelve research branches, including source separation, sound-event localization, binaural and Ambisonic rendering, spatial generation, neural room acoustics, HRTF personalization, loudspeaker sound-zone control, neural codecs and quality metrics.


### How many papers does the Spatial9 research review cover?

The review covers 124 academic records: the ninety-six records of the original research catalog plus twenty-eight additions on spatial reasoning, binaural enhancement, loudspeaker control, production interaction and differentiable processing. Four standards or assessment recommendations (ADM, IAMF, ITU-R BS.1534 and BS.1116) are discussed separately and not counted as academic models. The literature cutoff is 25 October 2026.


### What is Spatial9's main conclusion about AI spatial audio?

Spatial9's position is that creative intent must remain explicit, editable and testable from source to playback. No single research family owns the whole process; useful intelligence makes spatial decisions more manageable across the complete life of a scene, from separation and placement through revision, encoding and playback.


### Which source separation models are most relevant for immersive music production?

The review assesses Hybrid Transformer Demucs, Open-Unmix, Spleeter, Band-split RNN, BS-RoFormer, Mel-Band RoFormer, SAM Audio, AudioSep and MuS3D among others. Our verdict is that separation quality, inventory completeness and persistent source identity are separate judgments: a broad stem such as other may still need further decomposition before independent spatial placement makes sense.


### What is ACCDOA in sound event localization?

ACCDOA is an output representation for sound event localization and detection in which each class has a Cartesian vector: its length encodes whether the event is active and its orientation encodes the direction of arrival. The vector norm is neither source distance nor a calibrated probability, so activity detection and direction assignment should be decoded separately.


### Why does the FOA channel order matter in Ambisonics?

First-order Ambisonics can be stored in different channel orders and normalizations. Declaring the convention, for example ACN order W, Y, Z, X with SN3D normalization and a stated azimuth direction, makes channel mistakes observable before audio enters a neural model, a decoder or an export path. Models that use another convention need an explicit conversion.


### Can AI personalize HRTFs for binaural headphone listening?

Several reviewed methods, including HRTF Field, NIIRF and RANF, estimate listener-specific head-related transfer functions from limited information. Spatial9 separates profile estimation from playback: a renderer needs complete left and right filters with interaural timing, and filter-reconstruction error, front/back distinction, elevation and externalization should be reported as different outcomes.


### Is a single metric enough to measure spatial audio quality?

No. Fidelity, localization, externalization, immersion and artistic preference are related but distinct. Metrics such as SAQAM, AMBIQUAL, QASTAnet, SPAUQ and FAD each have their own validation domain, and listening protocols such as MUSHRA and BS.1116 answer specific questions. Spatial9 favors an interpretable profile of outcomes over one unexplained score.


### Are the code examples in the review official implementations?

Two examples, HTDemucs and SAM Audio, use official public project APIs checked against their sources. The other thirteen are original illustrative integration patterns that express Spatial9 architectural positions, including simplified translations of published equations. None are extracts from Spatial9 private implementation, and each lists its inputs, assumptions and omitted responsibilities.


## Sources

Numbers in square brackets throughout this article refer to this list. Publication status and the version reviewed are kept distinct; links point to the primary record used for the review.

- [1] An Audio-Visual System for Object-Based Audio: From Recording to Listening. 2018. IEEE Transactions on Multimedia 20(8), 1919–1931. Journal publication; metadata and associated dataset verified. Primary record
- [2] DDSP: Differentiable Digital Signal Processing. 2020. ICLR 2020. Conference paper; official research page links proceedings. Primary record
- [3] Automatic Multitrack Mixing with a Differentiable Mixing Console of Neural Audio Effects. 2021. IEEE ICASSP 2021; preprint first posted 2020. Conference publication. Primary record
- [4] Style Transfer of Audio Effects with Differentiable Signal Processing. 2022. Journal of the Audio Engineering Society 70(9), 708–721. Journal publication; DOI 10.17743/jaes.2022.0025. Primary record
- [5] Diff-MST: Differentiable Mixing Style Transfer. 2024. ISMIR 2024. Conference publication; author repository and demonstration manuscript confirm venue. Primary record
- [6] Diff-MSTC: A Mixing Style Transfer Prototype for Cubase. 2024. ISMIR 2024 demonstration. Demonstration manuscript; related CHI 2025 demo identified but not separately counted. Primary record
- [7] STASE: A spatialized text-to-audio synthesis engine for music generation. 2025. LLM4Music workshop at ISMIR 2025. Workshop acceptance stated in author manuscript. Primary record
- [8] SEND: A Spatial Event Neural Detector for Intentional Object Motion in Immersive Music Mixing. 2026. DAFx 2026. Conference proceedings. Primary record
- [9] Adoption of AI Technology in the Music Mixing Workflow: An Investigation. 2023. 154th Audio Engineering Society Convention. Convention research paper; author manuscript reviewed. Primary record
- [10] It's All About Speed: AI's Impact on Workflow in Music Production. 2025. AES International Conference on Machine Learning and Artificial Intelligence for Audio 2025; preprint posted 2026. Conference presentation stated in primary manuscript metadata. Primary record
- [11] Simon Rouard; Francisco Massa; Alexandre Défossez. Hybrid Transformers for Music Source Separation. 2023. ICASSP 2023. Peer-reviewed conference paper. Primary record
- [12] Fabian-Robert Stöter; Stefan Uhlich; Antoine Liutkus; Yuki Mitsufuji. Open-Unmix - A Reference Implementation for Music Source Separation. 2019. Journal of Open Source Software 4(41), 1667 (2019). Peer-reviewed software paper. Primary record
- [13] Romain Hennequin; Anis Khlif; Felix Voituret; Manuel Moussallam. Spleeter: a fast and efficient music source separation tool with pre-trained models. 2020. Journal of Open Source Software 5(50), 2154 (2020). Peer-reviewed software paper. Primary record
- [14] Yi Luo; Jianwei Yu. Music Source Separation with Band-split RNN. 2023. IEEE/ACM Transactions on Audio, Speech, and Language Processing 31, 1893–1901 (2023). Peer-reviewed journal paper. Primary record
- [15] Wei-Tsung Lu; Ju-Chiang Wang; Qiuqiang Kong; Yun-Ning Hung. Music Source Separation with Band-Split RoPE Transformer. 2024. ICASSP 2024. Peer-reviewed conference paper. Primary record
- [16] Ju-Chiang Wang; Wei-Tsung Lu; Minz Won. Mel-Band RoFormer for Music Source Separation. 2023. arXiv preprint; described as submitted to ISMIR 2023 late-breaking/demo track. Main-conference peer review not verified. Primary record
- [17] Bowen Shi et al. SAM Audio: Segment Anything in Audio. 2026. ICML 2026, Proceedings of Machine Learning Research 306, 112167–112194. Peer-reviewed conference paper. Primary record
- [18] Xubo Liu; Qiuqiang Kong; Yan Zhao; Haohe Liu; Yi Yuan; Yuzhuo Liu; Rui Xia; Yuxuan Wang; Mark D. Plumbley; Wenwu Wang. Separate Anything You Describe. 2025. IEEE Transactions on Audio, Speech and Language Processing 33, 458–471 (2025). Peer-reviewed journal paper. Primary record
- [19] Helin Wang et al. SAM Audio Judge: A Unified Multimodal Framework for Perceptual Evaluation of Audio Separation. 2026. arXiv:2601.19702, 27 January 2026. Preprint; peer-reviewed venue not verified. Primary record
- [20] V. V. Valtteri Kallinen; Eloi Moliner; Lauri Juvela; Vesa Välimäki. Music Source Separation via Stem Discovery. 2026. arXiv:2609.30912, 25 September 2026. Preprint; peer-reviewed venue not verified. Primary record
- [21] Bandhav Veluri; Justin Chan; Malek Itani; Tuochao Chen; Takuya Yoshioka; Shyamnath Gollakota. Real-Time Target Sound Extraction. 2023. ICASSP 2023. Peer-reviewed conference paper. Primary record
- [22] Sharath Adavanne; Archontis Politis; Joonas Nikunen; Tuomas Virtanen. Sound Event Localization and Detection of Overlapping Sources Using Convolutional Recurrent Neural Networks. 2019. IEEE Journal of Selected Topics in Signal Processing 13(1), 2019; early online 2018. Peer-reviewed journal paper. Primary record
- [23] Kazuki Shimada; Yuichiro Koyama; Naoya Takahashi; Shusuke Takahashi; Yuki Mitsufuji. ACCDOA: Activity-Coupled Cartesian Direction of Arrival Representation for Sound Event Localization and Detection. 2021. ICASSP 2021. Peer-reviewed conference paper. Primary record
- [24] Kazuki Shimada; Yuichiro Koyama; Shusuke Takahashi; Naoya Takahashi; Emiru Tsunoo; Yuki Mitsufuji. Multi-ACCDOA: Localizing and Detecting Overlapping Sounds from the Same Class with Auxiliary Duplicating Permutation Invariant Training. 2022. ICASSP 2022. Peer-reviewed conference paper. Primary record
- [25] Yin Cao; Turab Iqbal; Qiuqiang Kong; Fengyan An; Wenwu Wang; Mark D. Plumbley. An Improved Event-Independent Network for Polyphonic Sound Event Localization and Detection. 2021. ICASSP 2021. Peer-reviewed conference paper. Primary record
- [26] Kazuki Shimada et al. STARSS23: An Audio-Visual Dataset of Spatial Recordings of Real Scenes with Spatiotemporal Annotations of Sound Events. 2023. NeurIPS 2023 Datasets and Benchmarks. Peer-reviewed dataset paper. Primary record
- [27] David Diaz-Guerra et al. Baseline Models and Evaluation of Sound Event Localization and Detection with Distance Estimation in DCASE2024 Challenge. 2024. DCASE 2024 Workshop, 23–25 October 2024. Workshop proceedings paper. Primary record
- [28] Yusong Wu et al. Large-scale Contrastive Language-Audio Pretraining with Feature Fusion and Keyword-to-Caption Augmentation. 2023. ICASSP 2023. Peer-reviewed conference paper. Primary record
- [29] Bhavika Devnani; Skyler Seto; Zakaria Aldeneh; Alessandro Toso; Elena Menyaylenko; Barry-John Theobald; Jonathan Sheaffer; Miguel Sarabia. Learning Spatially-Aware Language and Audio Embeddings. 2024. NeurIPS 2024. Peer-reviewed conference paper. Primary record
- [30] Kentaro Seki; Yuki Okamoto; Kouei Yamaoka; Yuki Saito; Shinnosuke Takamichi; Hiroshi Saruwatari. Spatial-CLAP: Learning Spatially-Aware audio--text Embeddings for Multi-Source Conditions. 2026. ICASSP 2026. Peer-reviewed conference; official program verified. Primary record
- [31] Zhisheng Zheng; Puyuan Peng; Ziyang Ma; Xie Chen; Eunsol Choi; David Harwath. BAT: Learning to Reason about Spatial Sounds with Large Language Models. 2024. ICML 2024, PMLR 235:61454–61469. Peer-reviewed conference paper; later author revision reviewed. Primary record
- [32] Kevin Wilkinghoff; Zheng-Hua Tan. DSpAST: Disentangled Representations for Spatial Audio Reasoning with Large Language Models. 2026. ICASSP 2026. Peer-reviewed conference; official program verified. Primary record
- [33] Parthasaarathy Sudarsanam; Archontis Politis. Towards Spatial Audio Understanding via Question Answering. 2025. DCASE 2025 Workshop, 30–31 October 2025, Barcelona. Published workshop proceedings paper. Primary record
- [34] Zhiyuan Zhu et al. Spatial-Omni: Spatial Audio Understanding Integration in Multimodal LLMs via FOA Encoding. 2026. arXiv:2606.10738. Preprint; peer-reviewed venue not independently verified. Primary record
- [35] Jinbo Hu et al. MiDashengLM-Spatial: Unifying General Audio Understanding and Spatial Awareness. 2026. arXiv:2610.11156, first submitted 8 October 2026. Preprint; peer-reviewed venue not independently verified. Primary record
- [36] Yuhuan You; Lai Wei; Xihong Wu; Tianshu Qu. The World is Not Mono: Enabling Spatial Understanding in Large Audio-Language Models. 2026. arXiv:2601.02954. Preprint; peer-reviewed venue not verified through an accessible primary record. Primary record
- [37] Arvind Krishna Sridhar; Yinyi Guo; Erik Visser. Spatial Audio Motion Understanding and Reasoning. 2025. arXiv:2509.14666. Preprint; peer-reviewed venue not independently verified. Primary record
- [38] Artem Dementyev; Wazeer Zulfikar; Sinan Hersek; Pascal Getreuer; Anurag Kumar; Vivek Kumar. PhaseCoder: Microphone Geometry-Agnostic Spatial Audio Understanding for Multimodal LLMs. 2026. arXiv:2601.21124. Preprint; peer-reviewed venue not independently verified. Primary record
- [39] S Sakshi, Vaibhavi Lokegaonkar, Neil Zhang, Ramani Duraiswami, Sreyan Ghosh, Dinesh Manocha, Lie Lu. SPUR: A Plug-and-Play Framework for Integrating Spatial Audio Understanding and Reasoning into Large Audio-Language Models. 2025. arXiv:2511.06606. Preprint; no peer-reviewed venue identified in the inspected primary record. Primary record
- [40] Zhe Sun, Yujun Cai, Jiayu Yao, Yiwei Wang. AudioMotionBench: Evaluating Auditory Motion Perception in Audio LLMs. 2025. arXiv:2511.13273; revised title in January 2026. Preprint; no peer-reviewed venue identified in the inspected primary record. Primary record
- [41] Arvind Krishna Sridhar, Yinyi Guo, Erik Visser. Spatial Audio Question Answering and Reasoning on Dynamic Source Movements. 2026. arXiv:2602.16334. Preprint; no peer-reviewed venue identified in the inspected primary record. Primary record
- [42] Peiwen Sun; Sitong Cheng; Xiangtai Li; Zhen Ye; Huadai Liu; Honggang Zhang; Wei Xue; Yike Guo. Both Ears Wide Open: Towards Language-Driven Spatial Audio Generation. 2025. ICLR 2025. Peer-reviewed conference paper. Primary record
- [43] Mojtaba Heydari; Mehrez Souden; Bruno Conejo; Joshua Atkins. ImmerseDiffusion: A Generative Spatial Audio Latent Diffusion Model. 2025. ICASSP 2025. Peer-reviewed conference paper. Primary record
- [44] Saksham Singh Kushwaha; Jianbo Ma; Mark R. P. Thomas; Yapeng Tian; Avery Bruni. Diff-SAGe: End-to-End Spatial Audio Generation Using Diffusion Models. 2025. ICASSP 2025. Peer-reviewed conference paper. Primary record
- [45] Christian Templin; Yanda Zhu; Hao Wang. Generating Moving 3D Soundscapes with Latent Diffusion Models. 2026. ICASSP 2026. Peer-reviewed conference paper; IEEE record verified. Primary record
- [46] Zining Liang; Runbang Wang; Yujia Xiao; Xuzhou Ye; Qiuqiang Kong. ImmersiveFlow: Stereo-to-7.1.4 Spatial Audio Generation with Flow Matching. 2026. arXiv:2601.12950; latest reviewed v2 dated 28 September 2026. Preprint; named peer-reviewed venue not independently verified. Primary record
- [47] Ke Lei; Yu Zhang; Changhao Pan; Xueyi Pu; Wenxiang Guo; Ruiqi Li; Zhou Zhao. Towards Streaming Synchronized Spatial Audio Generation via Autoregressive Diffusion Transformer. 2026. arXiv:2605.30940, 29 May 2026. Preprint; peer-reviewed venue not verified. Primary record
- [48] Ziyu Luo; Lin Chen; Qiang Qu; Xiaoming Chen; Yiran Shen. DynFOA: Generating First-Order Ambisonics with Conditional Diffusion for Dynamic and Acoustically Complex 360-Degree Videos. 2026. arXiv:2602.06846; latest reviewed metadata v4, 23 July 2026. Preprint; peer-reviewed venue not verified. Primary record
- [49] Neural Synthesis of Binaural Speech from Mono Audio. 2021. ICLR 2021. Peer-reviewed conference. Primary record
- [50] BinauralGrad: A Two-Stage Conditional Diffusion Probabilistic Model for Binaural Audio Synthesis. 2022. NeurIPS 2022. Peer-reviewed conference. Primary record
- [51] Binaural Rendering of Ambisonic Signals by Neural Networks. 2022. arXiv 2211.02301. Preprint; later venue not verified. Primary record
- [52] Zero-Shot Mono-to-Binaural Speech Synthesis. 2025. Interspeech 2025. Peer-reviewed conference. Primary record
- [53] BinauralFlow: A Causal and Streamable Approach for High-Quality Binaural Speech Synthesis with Flow Matching Models. 2025. ICML 2025. Peer-reviewed conference. Primary record
- [54] Self-Supervised Generation of Spatial Audio for 360° Video. 2018. NeurIPS 2018. Peer-reviewed conference. Primary record
- [55] 2.5D Visual Sound. 2019. IEEE/CVF CVPR 2019. Peer-reviewed conference. Primary record
- [56] Sep-Stereo: Visually Guided Stereophonic Audio Generation by Associating Source Separation. 2020. ECCV 2020. Peer-reviewed conference. Primary record
- [57] Visually Informed Binaural Audio Generation without Binaural Audios. 2021. IEEE/CVF CVPR 2021. Peer-reviewed conference. Primary record
- [58] Geometry-Aware Multi-Task Learning for Binaural Audio Generation from Video. 2021. BMVC 2021. Peer-reviewed conference. Primary record
- [59] Cyclic Learning for Binaural Audio Generation and Localization. 2024. IEEE/CVF CVPR 2024. Peer-reviewed conference. Primary record
- [60] OmniAudio: Generating Spatial Audio from 360-Degree Video. 2025. ICML 2025. Peer-reviewed conference. Primary record
- [61] ViSAudio: End-to-End Video-Driven Binaural Spatial Audio Generation. 2025. arXiv 2512.03036. Preprint; later conference not verified. Primary record
- [62] Visually-Guided Spatial Audio Generation for 360° In-the-Wild Speech Scenes. 2026. Interspeech 2026. Accepted conference per author arXiv record. Primary record
- [63] PLACE: Positional Latent Adaptation via Conditioned Embeddings for Binaural Audio Generation. 2026. arXiv 2610.00630. Preprint. Primary record
- [64] Deep Learning Based Voice Extraction and Primary-Ambience Decomposition for Stereo to Surround Upmixing. 2023. AES 154th Convention Express Paper 62. Synopsis peer-reviewed; full express manuscript explicitly not peer-reviewed. Primary record
- [65] Automatic Audio Upmixing Based on Source Separation and Ambient Extraction Algorithms. 2023. IEEE SpeD 2023. Peer-reviewed conference. Primary record
- [66] A Deep Learning Approach to the Prediction of Time-Frequency Spatial Parameters for Use in Stereo Upmixing. 2024. DAFx 2024. Conference paper. Primary record
- [67] Ambisonizer: Neural Upmixing as Spherical Harmonics Generation. 2024. arXiv 2405.13428. Preprint; current record says under review. Primary record
- [68] DiffStereo: End-to-End Mono-to-Stereo Audio Generation with Diffusion Transformer. 2025. Interspeech 2025. Peer-reviewed conference. Primary record
- [69] S2X-Net: Stereo-to-Multichannel Neural Upmixing Network. 2026. IEEE Transactions on Audio, Speech and Language Processing 34:4362–4375. Peer-reviewed journal; accepted author manuscript reviewed. Primary record
- [70] Generative Learning for Ambisonic Upscaling. 2026. Accepted IEEE Transactions on Audio, Speech and Language Processing, 2026. Journal acceptance verified in author/university records; full author preprint reviewed. Primary record
- [71] Controllable Mono-to-Stereo Upmixing using Reward-tilted Flows. 2026. arXiv 2610.15967. Preprint. Primary record
- [72] Model-matching Principle Applied to the Design of an Array-based All-neural Binaural Rendering System for Audio Telepresence. 2023. IEEE ICASSP 2023. Peer-reviewed conference. Primary record
- [73] End-to-end multi-channel speaker extraction and binaural speech synthesis. 2025. arXiv 2410.05739. Preprint; later venue not verified. Primary record
- [74] Binaural Speech Enhancement Using Deep Complex Convolutional Transformer Networks. 2024. IEEE ICASSP 2024, pp. 681–685. Peer-reviewed conference. Primary record
- [75] Real-time multichannel deep speech enhancement in hearing aids: Comparing monaural and binaural processing in complex acoustic scenarios. 2024. IEEE/ACM Transactions on Audio, Speech, and Language Processing, DOI 10.1109/TASLP.2024.3473315. Peer-reviewed journal. Primary record
- [76] A Lightweight and Real-Time Binaural Speech Enhancement Model with Spatial Cues Preservation. 2025. IEEE ICASSP 2025, DOI 10.1109/ICASSP49660.2025.10888392. Peer-reviewed conference. Primary record
- [77] Interpretable Binaural Deep Beamforming Guided by Time-Varying Relative Transfer Function. 2026. arXiv 2511.10168, revised February 2026. Preprint; peer-reviewed venue not verified. Primary record
- [78] A Novel Binaural Cue Preservation Loss for DNN-Based Binaural Speech Enhancement. 2026. arXiv 2608.16299. Preprint; peer-reviewed venue not verified. Primary record
- [79] Learning Neural Acoustic Fields. 2022. NeurIPS 2022 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [80] SoundSpaces: Audio-Visual Navigation in 3D Environments. 2020. ECCV 2020 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [81] SoundSpaces 2.0: A Simulation Platform for Visual-Acoustic Learning. 2022. NeurIPS 2022 Datasets and Benchmarks peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [82] Visual Acoustic Matching. 2022. CVPR 2022 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [83] FAST-RIR: Fast neural diffuse room impulse response generator. 2022. IEEE ICASSP 2022 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [84] MESH2IR: Neural Acoustic Impulse Response Generator for Complex 3D Scenes. 2022. ACM Multimedia 2022 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [85] AV-NeRF: Learning Neural Fields for Real-World Audio-Visual Scene Synthesis. 2023. NeurIPS 2023 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [86] Neural Acoustic Context Field: Rendering Realistic Room Impulse Response With Neural Fields. 2023. arXiv 2023 and AV4D workshop copy; journal acceptance not verified. 2023 author preprint and AV4D workshop copy; journal acceptance not verified. Primary record
- [87] Real Acoustic Fields: An Audio-Visual Room Acoustics Dataset and Benchmark. 2024. CVPR 2024 peer-reviewed dataset/benchmark. Peer-reviewed publication, with reviewed author version identified. Primary record
- [88] Hearing Anything Anywhere. 2024. CVPR 2024 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [89] Direction-Aware Neural Acoustic Fields for Few-Shot Interpolation of Ambisonic Impulse Responses. 2025. Interspeech 2025 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [90] Physics-Informed Direction-Aware Neural Acoustic Fields. 2025. IEEE WASPAA 2025 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [91] Differentiable Room Acoustic Rendering with Multi-View Vision Priors. 2025. IEEE/CVF ICCV 2025 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [92] SOAF: Scene Occlusion-aware Neural Acoustic Field. 2026. IEEE/CVF WACV 2026 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [93] Multimodal Neural Acoustic Fields for Immersive Mixed Reality. 2025. IEEE Transactions on Visualization and Computer Graphics 31(5) peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [94] Explicit Context-Driven Neural Acoustic Modeling for High-Fidelity RIR Generation. 2026. Interspeech 2026 long paper, per author project and linked proceedings. Interspeech 2026 publication identified by author project; detailed review uses author preprint. Primary record
- [95] Few-shot Acoustic Synthesis with Multimodal Flow Matching. 2026. CVPR 2026 peer-reviewed; title verified in proceedings index. Peer-reviewed publication, with reviewed author version identified. Primary record
- [96] Image2Reverb: Cross-Modal Reverb Impulse Response Synthesis. 2021. IEEE/CVF ICCV 2021 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [97] Listen2Scene: Interactive material-aware binaural sound propagation for reconstructed 3D scenes. 2024. IEEE VR 2024. Peer-reviewed conference, accepted status verified from author primary sources. Primary record
- [98] HRTF Field: Unifying Measured HRTF Magnitude Representation with Neural Fields. 2023. IEEE ICASSP 2023 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [99] NIIRF: Neural IIR Filter Field for HRTF Upsampling and Personalization. 2024. IEEE ICASSP 2024 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [100] Retrieval-Augmented Neural Field for HRTF Upsampling and Personalization. 2025. IEEE ICASSP 2025 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [101] HRTF Interpolation using a Spherical Neural Process Meta-Learner. 2024. IEEE/ACM TASLP 32 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [102] Graph Neural Field with Spatial-Correlation Augmentation for HRTF Personalization. 2026. AAAI 2026 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [103] HRTF Upsampling and Personalization via Neural Operators. 2026. arXiv preprint; peer review not established. Preprint or workshop lineage; journal/conference acceptance not established beyond stated source. Primary record
- [104] HRTF Personalization via Sim-to-Real Neural Field. 2026. Interspeech 2026 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [105] SuDaField: Subject- and Dataset-Aware Neural Field for HRTF Modeling. 2025. IEEE Open Journal of Signal Processing 6 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [106] Towards HRTF Personalization using Denoising Diffusion Models. 2025. IEEE ICASSP 2025 peer-reviewed. Peer-reviewed publication, with reviewed author version identified. Primary record
- [107] SANN-PSZ: Spatially Adaptive Neural Network for Head-Tracked Personal Sound Zones. 2025. IEEE Transactions on Audio, Speech and Language Processing 33, 2735–2748. Peer-reviewed journal. Primary record
- [108] Neural personal sound zones with flexible bright zone control. 2025. arXiv author preprint. Preprint; primary conference acceptance not verified. Primary record
- [109] Synthesis of soundfields through irregular loudspeaker arrays based on convolutional neural networks. 2024. EURASIP Journal on Audio, Speech, and Music Processing 2024, article 17. Peer-reviewed journal. Primary record
- [110] Stereo Audio Rendering for Personal Sound Zones Using a Binaural Spatially Adaptive Neural Network (BSANN). 2026. IEEE Transactions on Audio, Speech and Language Processing, Early Access. Peer-reviewed journal early-access record from primary author institution. Primary record
- [111] Probing Spatial Structure in Pretrained Audio Representations. 2026. Interspeech 2026. Acceptance stated by authors and dataset record. Primary record
- [112] Can Large Language Models Understand Spatial Audio?. 2024. Interspeech 2024. Conference proceedings. Primary record
- [113] FOA Tokenizer: Low-bitrate Neural Codec for First Order Ambisonics with Spatial Consistency Loss. 2025. arXiv; submitted to ICASSP 2026. Preprint; acceptance not independently established. Primary record
- [114] SpatialCodec: Neural Spatial Speech Coding. 2024. IEEE ICASSP 2024; preprint 2023. Conference acceptance stated in author manuscript. Primary record
- [115] L3DAS22 Challenge: Learning 3D Audio Sources in a Real Office Environment. 2022. IEEE ICASSP 2022. Conference publication; DOI 10.1109/ICASSP43922.2022.9746872. Primary record
- [116] SAQAM: Spatial Audio Quality Assessment Metric. 2022. Interspeech 2022. Conference proceedings. Primary record
- [117] AMBIQUAL – a full reference objective quality metric for ambisonic spatial audio. 2018. QoMEX 2018. Conference publication. Primary record
- [118] AMBIQUAL: Towards a Quality Metric for Headphone Rendered Compressed Ambisonic Spatial Audio. 2020. Applied Sciences 10(9), 3188. Journal article; extension of AMBIQUAL. Primary record
- [119] QASTAnet: A DNN-based Quality Metric for Spatial Audio. 2026. IEEE ICASSP 2026; preprint first posted 2025. Conference DOI verified in arXiv metadata. Primary record
- [120] Quantifying Spatial Audio Quality Impairment. 2024. IEEE ICASSP 2024; preprint 2023. Conference proceedings. Primary record
- [121] Adapting Frechet Audio Distance for Generative Music Evaluation. 2024. IEEE ICASSP 2024. Conference publication. Primary record
- [122] SAVGBench: Benchmarking Spatially Aligned Audio-Video Generation. 2026. IEEE ICASSP 2026; first preprint 2024. Conference acceptance and author repository verified. Primary record
- [123] ASAudio: A Survey of Advanced Spatial Audio Research. 2025. IJCNLP-AACL 2025. Published survey. Primary record
- [124] Loss functions incorporating auditory spatial perception in deep learning -- a review. 2025. IEEE I3DA 2025. Conference DOI linked in author manuscript. Primary record
- [125] ITU-R BS.2076-3 Audio Definition Model. 2025. ITU-R Recommendation BS.2076-3. Standard; February 2025 revision. itu.int Primary record
- [126] Immersive Audio Model and Formats v2.0.0. 2026. Alliance for Open Media IAMF 2.0.0. Final deliverable dated 24 August 2026. aomediacodec.github.io Primary record
- [127] ITU-R BS.1534-3 Method for the subjective assessment of intermediate quality level of audio systems. 2015. ITU-R Recommendation BS.1534-3. Subjective assessment recommendation. itu.int Primary record
- [128] ITU-R BS.1116-3 Methods for the subjective assessment of small impairments in audio systems. 2015. ITU-R Recommendation BS.1116-3. Subjective assessment recommendation. itu.int Primary record
- [129] Spatial9 architecture demonstration. 2026. Spatial9 public product documentation. Company description; accessed 25 October 2026. spatial9.ai Primary record
- [130] Spatial9 features. 2026. Spatial9 public product documentation. Company description; accessed 25 October 2026. spatial9.ai Primary record
- [131] Spatial9 Acoustic Digital Twin demonstration. 2026. Spatial9 research preview. Company research preview; accessed 25 October 2026. spatial9.ai Primary record
- [132] Spatial9 ADM to Eclipsa conversion. 2026. Spatial9 public workflow documentation. Company description; accessed 25 October 2026. spatial9.ai Primary record
- [133] Spatial9 MCP demonstration. 2026. Spatial9 public workflow documentation. Company description; accessed 25 October 2026. spatial9.ai Primary record
- [134] BinauralSpeechSynthesis official repository and dataset conventions. Archived author repository; README dataset axes, tracking rate and audio rate inspected 25 October 2026. Primary record
- [135] IEM Plug-in Suite compatibility and AmbiX conventions. Official IEM documentation; ACN W Y Z X ordering, SN3D and unity W encoding gain inspected 25 October 2026. plugins.iem.at Primary record
- [136] PyTorch automatic differentiation with torch.autograd. Official autograd tutorial; requires_grad and backward usage inspected 25 October 2026. docs.pytorch.org Primary record
- [137] Demucs official Separator API implementation. Checked main on 2026-10-25: Separator constructor, separate_audio_file, samplerate and named-source dictionary. Repository is archived; no immutable commit SHA was exposed in the accessible source view. Primary record
- [138] Demucs official Python API documentation. Official repository documentation checked on 2026-10-25. Example is independently shortened from the documented public interface; inference was not executed. Primary record
- [139] SAM Audio official README: basic text prompting. Checked main on 2026-10-25 for checkpoint identifiers, processor inputs, CUDA usage and inference parameters. Gated checkpoint access is an installation assumption, not supplied by this excerpt. No immutable commit SHA was exposed. Primary record
- [140] SAM Audio official model.py: SAMAudio.separate. Checked main on 2026-10-25: separate accepts predict_spans and reranking_candidates; its return construction supplies target and residual as lists. This implementation detail takes precedence over the README tensor-style save example. Primary record
- [141] Official Learning Neural Acoustic Fields repository. Official entry point inspected; no repository API is claimed by this original adapter. Primary record
- [142] NumPy interp reference. Linear interpolation API and increasing-frequency requirement checked. numpy.org Primary record
- [143] NumPy convolve reference. Full linear convolution API checked. numpy.org Primary record
- [144] Official HRTF Field implementation. Magnitude-representation scope and SOFA data workflow checked. Primary record
- [145] SOFA SimpleFreeFieldHRIR convention. Data.IR, sampling rate and additional per-IR Data.Delay metadata checked. sofaconventions.org Primary record
- [146] NumPy rfft reference. Explicit transform length and axis API checked. numpy.org Primary record
- [147] NumPy irfft reference. Explicit inverse length, including odd lengths, checked. numpy.org Primary record
- [148] PyTorch einsum reference. Explicit frequency/point/speaker contraction checked. docs.pytorch.org Primary record
- [149] PyTorch functional ReLU reference. Nonnegative excess-gain penalty API checked. docs.pytorch.org Primary record
- [150] PyTorch complex-number documentation. Complex tensors and autograd guidance checked; runtime execution not available here. docs.pytorch.org Primary record