本文提出了基于3维空间Viterbi算法的汉语连续语音识别方法.本方法采用60个音素单位的隐马尔可夫模型(HMM)和8个声调单位的HMM作为识别用基元模型.音素基元模型和声调基元模型的识别结果的统合,采用音素单位的HMM状态、声调单位的HMM状态和时间的3维空间Viterbi算法来实现.语音声学处理和语音言语处理的结合,采用修改型Earley分析法的Top-Done型文法分析器和One Pass DP为基础的帧同步识别算法来实现.在由10名话者发音的有关旅馆预约指南的识别困难度是27.3的1070句子的识别实验中,总平均识别率达到94.4%.

This paper presents a recognition method of Chinese continuous speech,in which phonetic and prosodic features are integrated in terms of 3-Dimension Viterbi search.The phonetic information is modeled by 60 phonemic HMMs and the prosodic information by 8 tone HMMs.Both recognitions are synchronized based on 3-Dimension Viterbi search.A frame-synchronous parsing algorithm for CFG based on a top-down strategy is used for parsing processing.The task is related to the hotel reservation process,of which the perplexity is 27.3.For 1070 utterances produced by each of ten speakers,the average sentence recognition rate was 94.4%.