认人的限定主题的连续汉语语音识别系统的研究

Study on continuous Chinese speech recognition system for speaker-dependent and topic-constrained applications

本文描述一个基于矢量量化(VQ)、隐马尔可夫模型和有限态文法的认人的限定主题的连续汉语语音识别系统。引入跨零幅度差函数作为判定语音有无的特征参量之一,HMM训练用的各单个词语的语音数据由连续话句的语音数据经自动切分而得,识别过程中,每帧都考虑多个可能过渡到其它模型的文法节点。这些技术措施显著地提高了识别系统的准确率。这类系统能用于特定人操作的、特定主题的信息查询任务。待进一步解决非特定人的连续语音识别问题后,可用于特定主题的公用信息查询系统。

In this article a speaker-dependent, topic-constrained and continuous Chinese speech recognition system based on vector quantization (VQ), hidden Markov models (HMM's) and finite states syntactic analysis is studied. The difference of zero-crossing amplitudes was used as one of parameters to determine the position of voice start and end. The words-data used in HMM's training was obtained by machine segmentation of continuous spoken sentences. Multipossible transitions of models nodes were considere in each frame. So the recognition accuracy was improved. The system can be applied in speaker-dependent and topic constrained information request.