基于矢量量化和查找表的改进DTW语音识别方法

针对传统DTW语音识别方法的运算量和存储空间大的缺陷,提出一种基于矢量量化和查找表的改进DTW方法.方法利用矢量量化操作将连续特征矢量空间转化成离散矢量空间,以降低模式存储空间,在此基础上建立矢量失真测度表,并通过Hash查表方式实现了地址空间的精确定位,从而省去了动态规划操作造成的大量距离测度计算,极大提高了识别匹配速度.理论分析和实验结果证明了改进方法的有效性.同时为研究方便,在Matlab平台下设计和开发了DTW实时语音识别系统.

Improved DTW speech recognition method based on vector quantization and search table

In order to solve the disadvantages of traditional DTW speech recognition method with large computations and storages,an improved DTW based on vector quantization and search table is proposed.Firstly,the continuous feature vector space is translated into discrete form using vector quantization,with the purpose of reducing the model storage,and then the distortion table is built and accurate positioning of address space is realized by Hash search function,as a result,it can avoid lots of distortion computations cased by dynamic programming and largely increase the speed for recognition process.The theoretical analysis and experiment results prove that the im-proved method is effective.At last,a DTW based real-time recognition system under Matlab platform is developed.