Monthly Talk: Roberto Barumerli (Imperial College, UK)

Every month, the joint laboratory invites external speakers to take part in seminars for its partners.

Roberto Barumerli (Imperial College, UK): « Listening as an Inverse Problem: Bayesian Listener Models for Evaluating and Personalising Spatial Audio. »

Abstract:

Developing spatial audio systems for virtual and augmented realities relies on listening tests to verify perceptual outcomes. These tests are expensive, slow, and become the bottleneck in every design iteration. Auditory models that simulate how listeners localise sound using their head-related transfer functions (HRTFs) can instead make perceptual evaluation a tractable simulation problem. However, current models lack the statistical grounding needed to quantify listener behaviour across studies or generalise across experimental conditions.

In this talk, I will draw on computational neuroscience to frame localisation as an inverse problem solved via Bayesian inference. This approach yields a listener model that predicts how a given listener localises sound with any HRTF. I will show how the model accounts for individual differences and dynamic scenes, and how it serves as a perceptual evaluation engine for non-individual HRTF selection, HRTF upsampling and hearing protection devices. Its probabilistic formulation also supports statistically validated, model-based analysis, enabling competing models to be compared using statistical evidence. These results show how interpretable, uncertainty-aware listener models can provide principled perceptual metrics for data-driven spatial audio and reduce reliance on costly listening tests.

Bio:

Roberto Barumerli is a Marie Skłodowska-Curie Postdoctoral Fellow at the Dyson School of Design Engineering, Imperial College London, where he leads the MIA project on personalising spatial audio through computational modelling. He received his PhD in Information Engineering from the University of Padova, Italy, and was a postdoctoral researcher at the Acoustics Research Institute in Vienna and at the University of Verona, Italy. His research combines Bayesian modelling, statistics and listening experiments to build computational models of spatial hearing that serve both auditory science and the engineering of spatial audio technologies. He developed a Bayesian sound-localisation model, released as an open-source Python package, and co-organised the IEEE Listener Acoustic Personalisation (LAP) Challenge.