Auditory model-based similarity metric for head-related transfer functions
File(s)
Author(s)
Daugintis, Rapolas
Type
Thesis
Abstract
The head-related transfer function (HRTF) captures direction-dependent acoustic features shaped by the listener's anatomy that give rise to personal spatial auditory cues. While obtaining individual HRTFs at scale is impractical, studies on HRTF personalisation or user adaptation within binaural spatial audio applications require a meaningful metric to assess HRTF similarity. This is a challenging task due to the multimodal nature of auditory perception.
To address this, the thesis proposes an auditory model-based metric for evaluating non-individual HRTFs based on predicted localisation performance. The metric employs an existing Bayesian sound localisation model to compare non-individual HRTFs against the individual ones by matching their auditory cues. The thesis first details the model calibration and metric analysis using prior data from human localisation and HRTF rating studies, followed by perceptual validation through a series of listening tests.
A static sound localisation experiment demonstrates that the metric can reliably identify well-matched non-individual HRTFs that provide localisation performance similar to that of individual HRTFs, as well as poorly matched ones that significantly impair localisation. A dynamic spatial audio quality assessment further reveals that degraded localisation with non-individual HRTF correlates with perceived difference in overall rendering quality and tone colour, but not with externalisation or naturalness. Final studies evaluate the metric in a masked speech comprehension task using binaural rendering in reverberant conditions. While HRTF matching does not directly influence intelligibility, performance varies with reverberation rendering type and depends on acoustic HRTF features.
Overall, the validated similarity metric provides a perceptually grounded tool for assessing system-to-user and user-to-system HRTF adaptation strategies. However, perceptual evaluations suggest that the importance of HRTF personalisation depends on the listening scenario. These findings offer insights into potential optimisation of binaural audio personalisation methods.
To address this, the thesis proposes an auditory model-based metric for evaluating non-individual HRTFs based on predicted localisation performance. The metric employs an existing Bayesian sound localisation model to compare non-individual HRTFs against the individual ones by matching their auditory cues. The thesis first details the model calibration and metric analysis using prior data from human localisation and HRTF rating studies, followed by perceptual validation through a series of listening tests.
A static sound localisation experiment demonstrates that the metric can reliably identify well-matched non-individual HRTFs that provide localisation performance similar to that of individual HRTFs, as well as poorly matched ones that significantly impair localisation. A dynamic spatial audio quality assessment further reveals that degraded localisation with non-individual HRTF correlates with perceived difference in overall rendering quality and tone colour, but not with externalisation or naturalness. Final studies evaluate the metric in a masked speech comprehension task using binaural rendering in reverberant conditions. While HRTF matching does not directly influence intelligibility, performance varies with reverberation rendering type and depends on acoustic HRTF features.
Overall, the validated similarity metric provides a perceptually grounded tool for assessing system-to-user and user-to-system HRTF adaptation strategies. However, perceptual evaluations suggest that the importance of HRTF personalisation depends on the listening scenario. These findings offer insights into potential optimisation of binaural audio personalisation methods.
Version
Open Access
Date Issued
2025-09-30
Date Awarded
2026-03-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Picinali, Lorenzo
Geronazzo, Michele
Sponsor
European Union
Grant Number
101017743
Publisher Department
Dyson School of Design Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)