Interpretability of Neural Networks in Speech Processing

With the growing popularity of deep neural networks, the lack of transparency caused by their black box representation is raising demand for their interpretability. The goal of this thesis is to gain new insights into deep neural networks in speech processing tasks. Specifically, gender classification task on AudioMNIST dataset and speaker classification task on filterbanks from VoxCeleb dataset using convolutional and residual neural network. Layer-wise relevance propagation was used for the interpretation of these neural networks. This method produced heatmaps highlighting features that contributed positively and negatively to the correct classification. As results of interpretation show, classifications were mainly based on lower frequencies in time. In the case of gender classification, I managed to find the model's high dependency on a small number of features. Using obtained information, I created an augmented training set that increased the model's robustness.

deep neural networks, convolutional neural networks, speech processing, interpretation of neural networks, Layer-Wise Relevance Propagation