嵌入式语音识别系统中的DTW在线并行算法*

为提高语音识别系统的实时性，利用动态规划和并行计算思想，提出一种适用于嵌入式语音识别系统的DTW(动态时间规整)在线并行算法。通过分析标准DTW及其主要衍生算法，对DTW算法的数据结构进行改进以满足在线算法要求，在寻找最佳路径过程中动态连续地分配和释放内存或预先分配固定大小的内存，并将多个关键词的DTW计算分布到多个运算单元；最后汇总各运算单元的结果得到识别结果。实验表明，该算法比经典DTW降低了内存使用和识别时间，并使语音识别的实时系数达到1.17，具有较高的实时性。

Online parallel dynamic time warping algorithm for speech recognition in embedded system

The classical DTW can be enhanced using dynamic programming and parallel computing. This paper introduced an online parallel DTW to improve the real -time performance for embedded speech recognition systems. After comprehensive analysis of DTW and its major derivatives, the algorithm used data structures that fit the requirements of online algorithm. During the stage of figuring out optimal warping path, manipulated memory as dynamically allocated (and released) or statically allocated prior, and distributed calculations for each keyword to multiple computing units, then obtained the final recognition result from them. Experimental results indicate that the algorithm can reduce memory and time usage compared with classical DTW. Besides, the coefficient of real -time performances is reduced to 1.17, which is of high performance.