Deep neural networks as a model of speech perception

 It has been shown that when speech signals are presented to a person, they can be decoded from the electroencephalogram (EEG) using linear regression. Unfortunately due to the complex and nonlinear nature of the brain, the correlation between the actual and decoded signal are low and highly variable. In this project, we aim to improve this by leveraging deep learning architectures for automatic speech recognition. Firstly (1), we will build and improve an architecture to accurately classify EEG based on the presented stimulus. Next (2), we will link the accuracy of the previous architecture to the speech understanding level of the participants. Moreover, we will investigate which speech features are the most relevant to predict the speech understanding level. Finally (3), we will evaluate the deep neural network as a model for human speech perception. By investigating the effect of different speech features on the accuracy of the neural network, we can create a model of human speech recognition. These results have applications in auditory prostheses, and diagnostics of speech and language disorders. Next to the insight in neuroscience, we will also lay the foundation for brain decoders using deep learning techniques. Future applications are brain-computer interfaces that are able to directly read and act on people's thoughts.