语音识别片上系统中的多级搜索算法

Multi-Pass Decoding Algorithm Based on a Speech Recognition Chip

本文提出了一种新的用于片上的语音识别多级搜索算法 .该算法以连续隐含马尔可夫模型 (ContinuousDensityHMM ,CDHMM)为基本识别框架 .在保证识别率基本不变的前提下 ,大大降低了片内存储空间的占用量 ,减少了识别搜索时间 .在第二级识别候选词条的选取准则上 ,提出一种基于置信度的选择方法 ,更进一步改善了识别速度 ,增强了识别的稳健性 .在 2 0 0个语音命令的识别任务下 ,系统的识别率为 98.83% .而当识别词条增加到 6 0 0条时 ,该算法也具有良好的识别性能 .

A novel multi-pass decoding algorithm is proposed to realize a CDHMM-based embedded speech recognition system. Compared with traditional Viterbi decoding algorithm, both the memory consumption and recognition time decreases significantly with little recognition accuracy's reduction at the same time. Furthermore, the confidence measure is adopted as the principle to determine the candidates in the second pass, which makes the recognition process more efficient. For the task of 200 voice commands, the accuracy rate achieves 98.83%. And then, its performance can still keep steadily while the scale of recognition vocabulary is up to 600 commands.