Model-based Sparse Component Analysis for Multiparty Distant Speech Recognition

This research takes place in the general context of improving the performance of the Distant Speech Recognition (DSR) systems, tackling the reverberation and recognition of overlap speech. Perceptual modeling indicates that sparse representation exists in the auditory cortex. The present project thus builds upon the hypothesis that incorporating this information in DSR front-end processing could improve the speech recognition performance in realistic conditions including overlap and reverberation. More specifically, the goal of my PhD thesis is to exploit blind (source) separation of the speech components in a sparse space, also referred to as sparse component analysis (SCA), for multi-party multi-channel speech recognition.