One Pass Search Algorithm in Contnuous Speech Recognition System

This paper presents a new one pass search mode based on THSP(Tsinghua Speech) system. According to the characteristic of one pass mode, this paper improves THSP system in acoustic layer and language layer respectively. On one hand, in order to transmit the information of acoustic layer to language layer as more as possible, this paper improves the hypothesis of 3 choice of sentence length and the means of dealing with silence periods. On the other hand, this paper refers to the information of language layer to prune hopeless paths in acoustic layer. At last, according to the results of experiments, this paper compares one pass search mode based on viterbi search systems to two pass search mode and points out while the costs of each mode are at the same magnitude, the recognition rates of one pass search mode is obviously higher than those of two pass search mode.