Personalized HRTFs: Measurement, Modelling and ML

Module D: Advanced acoustics and rendering. Lesson 7 of 20, 30 min.

Abstract

Binaural renders of Atmos, IAMF and Ambisonic mixes are only as convincing as the HRTFs behind them. This lesson explains the acoustic cues HRTFs encode, why generic sets cause front-back and elevation errors, and how measurement, BEM simulation and machine learning are used to personalize them.

By the end you can

  • Explain how ITD, ILD and pinna spectral cues are encoded in HRIRs and why spectral cues are the least transferable between listeners
  • Load SOFA (AES69) HRIR files in Python and render a mono source binaurally at a chosen direction
  • Compare database selection, BEM simulation with Mesh2HRTF and machine-learning approaches to HRTF personalization
  • Design a physical, model-based and perceptual evaluation that exposes front-back and elevation errors

Every binaural render, from a headphone preview of a Dolby Atmos mix to an Eclipsa Audio stream decoded on a phone, passes through a set of head-related transfer functions. When those filters match the listener's own head and ears, virtual sources snap outside the head and stay put; when they do not, sounds collapse toward the centre of the head, flip from front to back, or drift in elevation. This lesson covers how HRTFs are measured, stored, simulated and personalized, and how to tell whether a personalization method actually works.

Why personalization matters

A head-related impulse response (HRIR) is the impulse response from a point in space to the entrance of one ear canal, measured or computed in free field (no room). Its Fourier transform is the head-related transfer function (HRTF). For a given direction, an HRTF set gives one filter per ear, and convolving a dry mono signal with that left/right pair places the sound in that direction over headphones. The physical effects encoded in the filters are diffraction around the head, reflections from the shoulders and torso, and the direction-dependent resonances and notches of the pinna.

Because heads and pinnae differ from person to person, so do HRTFs. The classic experiments by Wightman and Kistler (1989) showed that headphone stimuli synthesized with a listener's own HRTFs could produce localization judgements close to free-field listening. Wenzel and colleagues (1993) then asked what happens when listeners use a non-individual set taken from another person. Azimuth judgements stayed fairly robust, but front-back and up-down confusions became noticeably more frequent, and the effect varied a lot between listeners. That result still frames the field three decades later: generic HRTFs are good enough for many lateral cues, but elevation, externalization and front-back disambiguation depend on details that are personal.

For immersive music production this has practical consequences:

  • A binaural reference mix judged on one engineer's preferred HRTF may sound different to another listener, especially for height channels and overhead objects.
  • Height content in 5.1.4, Auro 3D or Atmos beds relies heavily on spectral (pinna) cues once it is rendered binaurally, which are the cues most damaged by mismatched HRTFs.
  • Head tracking reduces front-back confusions substantially because dynamic ITD changes disambiguate the cone of confusion, but it does not fully fix spectral mismatch for elevation.

For a perception refresher, see How We Hear in 3D. The rest of this lesson covers the acoustics of the cues, measurement and databases, the SOFA format, personalization by selection, simulation and machine learning, interpolation, and evaluation.

The rest of this lesson, its lab and self-check are part of Academy Advanced, a paid course. See who has access.

Key readings

  • Wightman, F. L., and Kistler, D. J. (1989). Headphone simulation of free-field listening. I: Stimulus synthesis. Journal of the Acoustical Society of America, 85(2).
  • Wenzel, E. M., Arruda, M., Kistler, D. J., and Wightman, F. L. (1993). Localization using nonindividualized head-related transfer functions. Journal of the Acoustical Society of America, 94(1).
  • Majdak, P., et al. (2013). Spatially Oriented Format for Acoustics: A data exchange format representing head-related transfer functions. Audio Engineering Society 134th Convention.
  • Brinkmann, F., Dinakaran, M., Pelzer, R., Grosche, P., Voss, D., and Weinzierl, S. (2019). A cross-evaluated database of measured and simulated HRTFs including 3D head meshes, anthropometric features, and headphone impulse responses. Journal of the Audio Engineering Society, 67(9).
  • Ziegelwanger, H., Kreuzer, W., and Majdak, P. (2015). Mesh2HRTF: An open-source software package for the numerical calculation of head-related transfer functions. 22nd International Congress on Sound and Vibration (ICSV22).