论文部分内容阅读
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。
This article describes a defined subject-based continuous Chinese speech recognition system based on vector quantization (VQ), hidden Markov models, and finite state grammars. The cross-zero amplitude difference function is introduced as one of the characteristic parameters to determine the presence or absence of speech. The speech data of each single word for HMM training is automatically segmented from the speech data of consecutive sentences. During the recognition process, Grammar nodes that may transition to other models. These technical measures significantly improve the accuracy of the recognition system. Such systems can be used for specific topic-specific information queries for specific human operations. To be further addressed non-specific continuous speech recognition problems, can be used for specific topics of public information query system.