Azimuth Sound Localization of Binaural Neural Networks Evaluating the biological plausibility and the exploitation of ILDs

Tichter, Alex

Azimuth Sound Localization of Binaural Neural Networks Evaluating the biological plausibility and the exploitation of ILDs

Files

Tichter, A. s-4579208-2021.pdf (3.44 MB)

Authors

Tichter, Alex

Issue Date

2021-04-01

Language

en

URI

https://theses.ubn.ru.nl/handle/123456789/16157

Abstract

Normal-hearing humans can accurately localize sounds and use this ability to focus on a speci c sound source in noisy environments. This accuracy can only be achieved through the processing of acoustic cues in our brain. Recently, it has been shown that also arti cial neural networks are able to mimic human sound localization performance under various listening conditions. However, it is still unknown whether the arti cial networks process the same cues in a biologically plausible manner. In addition to the commonly used broadband, high-pass and low-pass stimulus- response plots for sound localization, this master thesis analyzed the spatial, fre- quency and ILD tuning of arti cial binaural neural networks with varying com- plexity. The results suggest that ILDs contribute to, but can not fully explain the localization strategy of the networks. Especially narrow frequency tuning and an insensitivity to low-pass sounds was absent in all tested models. Based on these results, we hypothesize that a more complex input, such as complex sounds, and the addition of a new learning goal (i.e. pitch detection or musical categorization) is needed to facilitate the development of biologically-plausible arti cial neural networks.