连续语音识别中利用帧间相关性的研究

Research on Exploiting Inter－Frame Correlation in Continuous Speech Recognition

尽管作为当前最为流行的语音识别模型，隐马尔可夫模型（HMM）由于采用了状态输出独立同分布假设，因此不能描述语音现象中固有的时间相关性。文章介绍了一个更为灵活的基于段长分布HMM（DDBHMM）的研究帧相关性的框架，并在此基础上提出了一个混合模型，采用一种将语音特征静态信息和动态变化信息分别描述又有机结合在一起的方式，以较小的计算代价更为合理地刻划了真实的语音现象。汉语大词汇量非特定人连续语音识别的实验表明，通过利用帧相关性识别系统的性能得到了明显改善。

： Although as the most popular model for speech recognition，HMM assumes the output observations of a state are independent and identically distributed，thus is unable to describe the temporal correlation inherent in speech phenomena．This paper introduces a more flexible framework for exploiting inter－frame correlation based on Duration Distribution Based Hidden Markov Model（DDBHMM）．Under the above framework we propose a new model which describes the real speech phenomena more reasonably at low computation cost，by modeling separately the static and dynamic characteristics of speech features and combing them into an integrated model．The experiments for Chinese large－vocabulary speaker－independent continuous speech recognition show that exploiting the correlation improves the performance of the recognition system distinctively．