Conference Paper
Learnable Gabor filters for human auditory localisation
AES Show 2026 Nashville · Paper 25 · October 2026
Abstract
This paper investigates a novel approach to biologically inspired modelling of human sound source localisation, whereby the spectro-temporal receptive fields of auditory neurons, represented as parametrised and learnable Gabor filters, are used as an optimisable mid-brain representation. While previous research treats the mid-brain as either a learnable black box or a non-learnable bio-inspired representation, this work proposes a learnable mid-brain representation that is explainable by its relation to an existing model of simple-cell responses. The use of parametrised filters reduces the number of learned parameters compared with traditional convolutional kernels whilst re-formulating the convolutional layers as Fourier domain operations allows for a reduction in computation
time. The performance of the model is compared against existing perceptual data to evaluate model-human alignment and is also compared to other, comparable, recent models from the literature. Results showed that whilst
the proposed model performs to the level beyond what would be expected of a human listener, it does exhibit patterns associated with human behaviour, such as front-back confusions. A preliminary single-seed ablation that removes the attention layer showed potential benefit in bringing the model closer to alignment with human performance, however further investigation is required.
