Speech Recognition Based on Intelligent Beam Search Algorithm

在大词汇量的连续语音识别中，有效的识别算法的研究是非常重要的，因为匹配过程需要在现实的CPU时间内执行巨大的计算量。传统的基于动态时间规整(DTW)、隐马尔可夫模型(HMM)等的识别算法是在穷举搜索可能的组合的基础上构建的。在基于DTW和HMM的语音识别系统中，采用动态规划技术进行穷举搜索，将匹配过程看作是通过一个约束节点搜索最优路径的问题。在图搜索算法在语音识别中的应用中，有两种搜索算法是有效的，即波束搜索算法和最佳优先搜索算法。在使用波束搜索算法的语音识别中的传统剪枝策略仅基于从开始节点到当前节点的分数。不使用从当前节点到终端节点的分数估计。将A^*算法引入到基于最佳优先搜索算法的语音识别中，从图搜索的角度对匹配过程进行建模，描述了基于DTW和HMM的语音识别算法的新方法。第一章描述了一种基于波束搜索算法的基于DTW的语音识别。第二章介绍了一种基于最佳优先搜索算法的基于DTW的语音识别方法。最后，在第三章中，描述了一种利用最佳优先搜索算法的HMM语音识别方法。

In a large-vocabulary continuous speech recognition, an investigation of efficient recognition algorithms is extremely important because of executing an enormous computation needed in a matching process within a realistic CPU time. Conventional recognition algorithms based on a dynamic time warping (DTW), a hidden Markov model (HMM) and so on are constructed on the base of an exhaustive search of possible combinations. A dynamic programming technique is introduced to execute the exhaustive search efficiently.The matching process in DTW-based and HMM-based speech recognition systems is regarded as a problem of searching an optimal path through a constrained node. In an application of graph searching algorithms to speechrecognition, two kinds of searching algorithms are effective, that is, a beam searching algorithm and a best-first searching algorithm. A conventional pruning strategy in speech recognition using the beam searching algorithm is based on only a score from the beginning node to the current node. A score estimate from the current node to the terminal node is not used. An A^* algorithm is introduced to speech recognition using the best-first searching algorithm.This report describes new approaches to DTW-based and HMM-based speech recognition algorithms by modeling the matching process from a view point of a graph search. In Chapter I, a DTW-based speech recognition utilizing the beam searching algorithm is described. In Chapter II, a DTW-based speech recognition utilizing the best-first searching algorithm is described. Finally in Chapter III, an HMM-based speech recognition utilizing the best-first searching algorithm is described.