No abstract

Our ability to comprehend speech remains, to date, unrivaled by deep learning models. This feat could result from the brain’s ability to fine-tune generic sound representations for speech-specific processes. To test this hypothesis, we compare i) five types of deep neural networks to ii) human brain responses elicited by spoken sentences and recorded in 102 Dutch subjects using functional Magnetic Resonance Imaging (fMRI). Each network was either trained on an acoustics scene classification, a speech-to-text task (based on Bengali, English, or Dutch), or not trained. The similarity between each model and the brain is assessed by correlating their respective activations after an optimal linear projection. The differences in brain-similarity across networks revealed three main results. First, speech representations in the brain can be accounted for by random deep networks. Second, learning to classify acoustic scenes leads deep nets to increase their brain similarity. Third, learning to process phonetically-related speech inputs (i.e., Dutch vs English) leads deep nets to reach higher levels of brain-similarity than learning to process phonetically-distant speech inputs (i.e. Dutch vs Bengali). Together, these results suggest that the human brain fine-tunes its heavily-trained auditory hierarchy to learn to process speech.

Speech classifiers of paralinguistic traits traditionally learn from diverse hand-crafted low-level features, by selecting the relevant information for the task at hand. We explore an alternative to this selection, by learning jointly the classifier, and the feature extraction. Recent work on speech recognition has shown improved performance over speech features by learning from the waveform. We extend this approach to paralinguistic classification and propose a neural network that can learn a filterbank, a normalization factor and a compression power from the raw speech, jointly with the rest of the architecture. We apply this model to dysarthria detection from sentence-level audio recordings. Starting from a strong attention-based baseline on which mel-filterbanks outperform standard low-level descriptors, we show that learning the filters or the normalization and compression improves over fixed features by 10% absolute accuracy. We also observe a gain over OpenSmile features by learning jointly the feature extraction, the normalization, and the compression factor with the architecture. This constitutes a first attempt at learning jointly all these operations from raw audio for a speech classification task.