Â鶹´«Ã½

Title of the thesis: Learning Deep Acoustic Feature Representations for HRTF Individualization

Thesis defender: Etienne Thuillier
Opponent: Dr. Fabian Brinkmann, TU Berlin, Germany 
Custos: Prof. Vesa Välimäki, Aalto University School of Electrical Engineering 

The growing adoption of augmented and virtual reality technologies requires immersive spatial audio that can scale to consumer devices. A key component is the head-related transfer function (HRTF), which describes how sound interacts with a listener’s head, ears, and torso, depending on the direction of the sound source. Because HRTFs depend on individual morphology, using generic approximations degrades spatial perception. Unfortunately, measuring individualized HRTFs requires specialized facilities and lengthy procedures, making large-scale deployment impractical.

Deep learning provides a scalable, data-driven alternative for HRTF individualization. This thesis identifies key challenges in predicting individualized HRTF from data and proposes three complementary solutions.

First, to better align predictive models with human perception while reducing reliance on costly listening tests, explainable AI methods are used to reveal salient spectral cues that are relevant for sound localization. Second, a probabilistic, geometry-aware neural model accurately estimates HRTFs at unmeasured directions with reliable uncertainty, reducing the number of required measurements by up to a factor of two while enabling adaptive acquisition strategies. Third, a generative diffusion model reconstructs personalized HRTFs from reverberant binaural recordings using consumer-grade microphones, enabling estimation in everyday environments without specialized facilities.

Together, these contributions advance the scalable personalization of spatial audio, demonstrating that individualized sound rendering can be achieved with reduced data and in realistic conditions. This supports the deployment of more immersive audio experiences in everyday devices such as headphones and augmented or virtual reality systems.

Thesis available for public display 7 days prior to the defence at .

Contact: etienne.thuillier@aalto.fi