正反向隐马尔可夫模型及其在连续语音识别中的应用

本文针对语音信号中客观存在的正、反向依赖特性，明确提出了用条件概率的概念来定量表述语音信号的这种正、反向的马尔可大依赖关系，提出了描述语音信号这种正反向依赖关系的正反向隐马尔可夫模型（ＨＭＭ），并用实验证明了仅仅利用语音反向依赖关系语音识别同样也能获得相当可观的识别性能。接着，本文针对孤立字和连续语音两种不同的识别任务，研究了在语音识别中同时利用这两种依赖信息的方法，并提出了一种连续语音识别中的新的搜索算法──正反向分半混合搜索。这种方法利用基于正向ＨＭＭ的正向Ｖｉｔｅｒｂｉ搜索和基于反向ＨＭＭ的反向Ｖｉｔｅｒｂｉ搜索的中间结果来有效地结合正反向依赖信息，实验证明正反向分半混合搜索方法确实一致地优于单用任何一种依赖信息的单向搜索识别方法。

Forward and Backward Hidden Markov Model with Their Applications to Continuous Speech Recognition

In view of objectively-existing forward and backward contextual dependent information in speech signal, in this paper we show that a conditional probability can be used to explicitly express these information in speech signal, which is called Forward and Backward Markov Contextual Dependences. And the Forward and Backward Hidden Markov Models, which are believed to contain these two Markov Contextual Dependences respectively, are also presented here. Our experimental results prove that we can obtain the relatively acceptable performance in speech recognition only using the Backward Markov Contextual Dependence. It is at least comparable with the recognition performance obtained from normal forward HMM with Forward Markov Contextual Dependence. In this paper, we also study the methods of simutaneously taking advantages of these two Markov Contextual Dependences of speech in isolated word recognition and continuous speech recognition respectively, and propose an effective algorithm in continuous speech recognition-Forward and Backward Mixed Bisected Search(FBMBS), which can utilize the immediate results of Forward HMM based forward Viterbi search and Backward HMM based backward Viterbi search,so that we can conveniently combine these two Markov Contextual Dependences in continuous speech recognition at the same time. Our experiments show that the FBMBS algorithm really is consistently superior to that of using either of these two Contextual Dependences alone.