HMM非特定人连续语音识别的嵌入式实现

Embedded Implementation of HMM Speaker-Independent Continuous Speech Recognition System

嵌入式系统正逐渐成为语音识别实际应用的首选平台。该文在嵌入式平台上研究HMM连续语音识别的计算复杂度要素,提出特征系数屏蔽方法和综合剪枝相结合的'瘦身'计算方法,降低计算复杂度并保持识别率。该方法在嵌入式平台上研究的实验数据表明,HMM连续语音识别瘦身系统与基线系统相比,计算时间从基线系统的100％降低到27.91％,识别率仅从基线系统的89.65％下降到89.41％。

The embedded systems are gradually becoming the first choice of platforms which should be used for real-time speech recognition system. This paper discusses the computation complexity factors of HMM-based continuous speech recognition for embedded system. An optimized way integrating feature masking and pruning is presented to reduce the computation complexity and keep the recognition accuracy. The experiments for embedded system show that, comparing with the base-line system, the computation time is reduced from 100% to 27.91%, and the recognition accuracy is degraded only from 89.65% to 89.41%.