Towards phonetic interpretability in deep learning applied to voice comparison


A deep convolutional neural network was trained to classify 45 speakers based on spectrograms of their productions of the French vowel /ɑ̃/ Although the model achieved fairly high accuracy – over 85 % – our primary focus here was phonetic interpretability rather than sheer performance. In order to better understand what kind of representations were learned by the model, i) several versions of the model were trained and tested with low-pass filtered spectrograms with a varying cut-off frequency and ii) classification was also performed with masked frequency bands. The resulting decline in accuracy was utilized to spot relevant frequencies for speaker classification and voice comparison, and to produce phonetically interpretable visualizations.

International Congress of Phonetic Sciences