Quantitative models of hearing and perception
4 peer-reviewed publications · 1 PhD thesis
Deriving binaural computations with deep learning
University of Nottingham · Hearing Sciences, UK · Advisor: Prof. Michael Akeroyd
Binaural detection is how the brain uses spatial information to detect sounds masked by background noise. Deep neural networks were trained to reproduce this behavior, then interrogated to reverse-engineer the mechanism in silico, converging on the same computation the brain is thought to use.
Methods: variational autoencoders · explainable AI · TensorFlow · Python · closed-form equation derivation
Inferring the basis of binaural detection with a modified autoencoder. Frontiers in Neuroscience 17:1000079 (2023).
Trained autoencoder networks to reproduce human binaural detection behaviour, then inspected the trained networks directly: the optimal networks had spontaneously developed timing-sensitive units consistent with a cross-correlation mechanism, reproducing classic human masking-level-difference data.
Paper: Frontiers in Neuroscience · Code: GitHub
A set of equations for numerically calculating the interaural level difference in the horizontal plane. JASA Express Letters 1(4), 044402 (2021).
Derived closed-form equations for computing interaural level difference at any frequency and horizontal-plane angle, fit to classic head-related transfer function measurements to within 0.5 dB.
Paper: JASA Express Letters
Equations for multi-stage audiovisual integration
Nottingham Trent University, UK · Advisor: Dr. Christian Sumner
Watching someone's face helps you understand speech in noisy environments, but whether that benefit comes from earlier sensory processing or later integration is disputed. This work derived equations to examine the question directly.
Methods: signal detection theory · mathematical derivation · Bayesian modelling · R
Equations unifying early- and late-noise accounts of audiovisual integration.
This work derived equations, from first principles, describing the relative contributions of early (sensory) and late (decision-stage) noise to audiovisual performance, using signal detection theory within a Bayesian framework. This project is ongoing, and data is now being collected to fit and test the model directly.
Presented at Speech in Noise Workshop 2025, ARO 2023.
Machine-learned phenotypes of tinnitus
University of Nottingham · Hearing Sciences, UK · Collaboration: Nottingham Health Informatics Collective
Identifying tinnitus phenotypes matters for individualizing treatment, but earlier attempts using clustering algorithms have cast doubt on whether such phenotypes even exist. This work used a probabilistic model that avoids forcing patients into fixed clusters, and found four underlying phenotypes, each linked differently to patients' tinnitus.
Methods: unsupervised learning · psychological questionnaires
An exploration of psychological symptom-based phenotyping of adult cochlear implant users with and without tinnitus using a machine learning approach. Progress in Brain Research 260, 283–300 (2021).
Used Gaussian mixture modelling and cross-validation to test whether adult cochlear implant users with tinnitus form distinct psychological symptom-based phenotypes (anxiety, depression, insomnia), addressing conflicting evidence from earlier tinnitus-phenotyping studies that relied on simpler clustering methods.
Peer-reviewed book chapter: Progress in Brain Research
Presented at BCIG 2020.
Inferring speech-in-noise from neural codes
University of Nottingham · Hearing Sciences, UK · Advisors: Dr. Christian Sumner, Prof. Michael Akeroyd, Dr. Mark Wallace
Classical accounts assume the brain segregates competing sounds using cues like pitch and timing before recognition happens. This work modelled recognition directly from midbrain neural codes, using a classifier that infers the whole scene at once. It reproduced human performance and errors without ever needing to extract those cues.
Methods: multi-channel electrophysiology · auditory scene analysis · ideal-observer modelling · naive Bayes classification · speech synthesis
Neural coding of overlapping speech and the effects of hearing loss. PhD thesis (Otorhinolaryngology), University of Nottingham, UK (2020).
Proposed a model of concurrent vowel perception that, unlike standard segregation-based accounts, predicts the combined neural representation of overlapping vowels without first separating them, matching both midbrain recordings and human perceptual data, and extending to show how hearing loss degrades recognition of overlapping speech.
Thesis: Nottingham ePrints
Revisiting models of concurrent vowel identification: the critical case of no pitch differences. Acta Acustica united with Acustica 104(5), 922–925 (2018).
Built a computational model that predicts listeners' vowel-pair confusions when two vowels share the same pitch (a case prior segregation-based models badly under-predicted) by modelling the combined neural representation directly rather than separating the two vowels first.
Paper: Acta Acustica
Medical Research Council Studentship, £60,000 (2016–2019).
Presented at Computational Audiology 2020, CNS 2019, SfN 2018.