To Separate Speech! A System for Recognizing Simultaneous Speech

The PASCAL Speech Separation Challenge (SSC) is based on a corpus of sentences from the Wall Street Journal task read by two speakers simultaneously and captured with two circular eight-channel microphone arrays. This work describes our system for the recognition of such simultaneous speech. Our system has four principal components: A person tracker returns the locations of both active speakers, as well as segmentation information for each utterance, which are often of unequal length; two beamformers in generalized sidelobe canceller (GSC) configuration separate the simultaneous speech by setting their active weight vectors according to a minimum mutual information (MMI) criterion; a postfilter and binary mask operating on the outputs of the beamformers further enhance the separated speech; and finally an automatic speech recognition (ASR) engine based on a weighted finite-state transducer (WFST) returns the most likely word hypotheses for the separated streams. In addition to optimizing each of these components, we investigated the effect of the filter bank design used to perform subband analysis and synthesis during beamforming. On the SSC development data, our system achieved a word error rate of 39.6%.