嵌入式英语命令词语音识别算法研究  

Embedded English Speech Commands Recognition Algorithm Research

作　　者：姚竞[1] 王国梁[1] 刘加[1] 

机构地区：[1]北京清华大学电子工程系,100084

出　　处：《微计算机信息》2008年第17期4-5,8,共3页Control & Automation

基　　金：国家自然生科学基金资助项目(60572083)

摘　　要：本文提出了一种基于定点DSP的嵌入式英语语音命令词识别算法,并基于TI芯片建立识别系统。系统采用基于连续隐Markov模型(Continuous Density Hidden Markov Model,CDHMM)的两阶段识别策略。通过决策树结合数据驱动的状态聚类方法,一阶段模型数目研究等方法提高识别率。最后在以TI TMS320vc5502定点DSP为核心的语音处理片上系统上实现了英语语音命令识别,当DSP工作速度为200MIPs时,实时率为0.37,存储空间消耗为49.5kbyte,对于1235词的识别效果为95.4%。This paper presents an embedded English commands speech recognition system based on fixed-point DSP and establishes recognition system based on TI chip. The system uses Continuous Density Hidden Markov Model （CDHMM） and two-pass search strategy. To increase the recognition accuracy, phone models tying based on decision tree and data driven, phone model unit choosing in first stage recognition are applied in the system. The English commands speech recognition system has been realized on speech processing system on chip based on TI TMS320vc5502 fixed point DSP. When the DSP works on speed of 200MIPs, the system can provide a recognition accuracy rate of 95.4 % on 1235 two-word phrases recognition task with 0.46 times lower than real time and 49.5kbyte data space.

关 键 词：语音识别 嵌入式 状态共享 特征选择 

分 类 号：TN912.3[电子电信—通信与信息系统]