Integrating induced probability into decoding for large vocabulary continuous speech recognition

This paper integrates location information of frames into conventional acoustic model (AM) and language model (LM) likelihoods, in order to distinguish potential path candidates more precisely at decoding stage. This paper proposes an induced probability, which represents location information of frames within the whole acoustic space. By integrating the induced probability, the decoder is directed to search within the most promising regions of acoustic space. Promising paths are enhanced and unlikely paths are weakened. Experiments conducted on Chinese Putonghua show that the character error rate is reduced by 10.95% relatively without increasing decoding complexity significantly. Finally, pruning analysis shows that integrating location information of frames into traditional decoding framework is helpful for improving system performance.