基于DHMM的嵌入式语音识别系统的实现与优化

Implementation and Optimization of Embedded Speech Recognition System Based on DHMM

为满足在嵌入式系统上实现非特定人的语音识别需求,介绍一种采用离散隐马尔科夫模型(DHMM)的嵌入式语音识别系统的实现方法。设计出一种在嵌入式系统上完成语音识别功能、在PC系统上完成模板训练功能的软件架构。针对TMS320VC5509A处理器的特点,对识别软件进行实现和优化。根据语音信号不可逆的特点,提出一种优化的viterbi算法。实验结果表明,在保证识别率的条件下系统取得了200 ms以内的实时响应速度,具有很高的实用性。

In order to meet the demand of realizing the speaker-independent speech recognition on the embedded system, a embedded speech recognition system in discrete hidden Markov model （DHMM） is proposed. The software architecture of the recognition software part rtmning on embedded system and the training software part running on PC system are designed. According to features of TMS320VC5509A processor, the recognition software on embedded system is implemented and optimized. Also it is presented that the Viterbi algorithm is optimized on the basis of the irreversible characteristics of the speech signal. Experiments show the real-time response speed of the system achieved within 200 ms at same the recognition rate, and show its high availability.