Bio-inspired neural network for human auditory localisation

Abstract

This study presents a lightweight, biologically inspired model of human sound localisation in which interaural time and level differences (ITD and ILD) are encoded as excitation-inhibition (EI) patterns, emulating binaural processing in the auditory brainstem. A tonotopically partitioned multilayer perceptron (MLP) processes these patterns by analysing neighbouring frequency channels in parallel and integrating them progressively across broader bands, mirroring the cochlea-to-cortex frequency mapping. This inductive bias guides learning toward biologically plausible solutions with fewer parameters than a fully connected design. Model-human alignment is validated against a controlled anechoic localisation experiment using burst noise, enabling quantitative comparison of full response distributions. The model closely matches human listeners in both accuracy and error patterns. Additional psychophysical simulations test whether it reproduces broader, multidimensional localisation behaviours. The model reproduces key human localisation behaviours: higher frequencies reduce front-back confusions in speech; cue weighting shifts from ITD to ILD dominance with increasing frequency; and accuracy improves with bandwidth. To identify which frequency partitions drive MLP decisions, each partition was systematically zeroed at inference. The learned strategy proves frequency-specific and stimulus-dependent: reliance shifts from low to high frequency partitions as spectral content increases, with front-back disambiguation concentrating in the highest bands. Together, these findings demonstrate that physiologically motivated cue embeddings combined with frequency-organised learning yield a compact, interpretable, and human-aligned model of azimuth perception.

More information

Identifiers

ORCID for Hsuan-Yang Wang: orcid.org/0000-0002-0890-179X

ORCID for Philip Nelson: orcid.org/0000-0002-9563-3235

ORCID for Christine Evers: orcid.org/0000-0003-0757-5504

Catalogue record

Export record

Altmetrics