基于3维空间Viterbi算法的音素模型和声调模型识别概率统合方法的研究

Study on the integration of phonetic and prosodic probability based on 3-dimension viterbi search

提出了一种在汉语连续语音识别中基于 ３维空间 Ｖｉｔｅｒｂｉ算法的音素模型和声调模型识别概率的统合方法。该方法采用６０个音素单位的ＨＭＭ和８个声调单位的ＨＭＭ作为识别用基元模型。音素和声调基元模型识别结果的统合，采用音素的ＨＭＭ状态、声调的ＨＭＭ状态和时间的３ 维空间帧同步Ｖｉｔｅｒｂｉ 算法来实现。本文还探讨了在该方法的基础上，给予不同路径限制时的匹配统合效果，并且通过和传统的匹配统合方式的比较，证明了提出的方法的有效性。

This paper presents a new method of continuous speech recognition for Chinese, in which phonetic and prosodic features were integrated in terms of 3-Dimension Viterbi search.The phonetic information was modeled as 60 phonemic HMMs and 11 tone HMMs of the prosodic information. Both models are synchronized based on 3-Dimension Viterbi search. We investigated integration methods of phonetic and prosodic likelihoods based on different at search paths and compared them with traditional method through the experiments on continuous speech recognition of Chinese. The efficiency of the proposed approach is verified in this paper.