An improved steady segment based decoding algorithm by using response probability for LVCSR

Abstract

This paper proposes a novel decoding algorithm by integrating both steady speech segments and observations' location information into conventional path extension framework. First, speech segments which possess stable spectrum are extracted. Second, a preliminarily improved algorithm is given by modifying traditional inter-HMM extension framework using the detected steady segments. Then, at probability calculation stage, response probability (RP), which represents location information of observations within acoustic feature space, is further incorporated into decoding. Thus, RP directs the decoder to enhance/weaken path candidates that get through the front end steady-segment-based decoding. Experiments conducted on Mandarin speech recognition show that character error rate of proposed algorithm achieves a 4.6% relative reduction when compared with a system in which only steady segment is used, and run time factor achieves a 10.0% relative reduction when compared with a system in which only RP is used.