HMM非特定人孤立词语音识别系统的片上实现

On Chip Realization of HMM Speaker-independent Isolated Word Speech Recognizer

在SEED-DEC5502DSP嵌入式系统开发平台上实现了一个面向非特定人的孤立词语音识别系统,与传统的基于特定人的语音识别系统相比,该系统无需用户训练,易于使用。系统采用改进的基于语音对数域能量变化率的实时端点检测算法,仅对检测的有声段语音进行特征提取和解码,减少了要处理的语音帧数;对状态输出概率计算进行了分析和优化,进一步降低了计算负担。实验表明系统在100词条的情况下识别率达到98%,识别时间为1.03倍实时。

An embedded speaker-independent isolated word speech recognition system is designed and realized in the SEED-DEC5502 EVM platform. Compared with the speaker-dependent system, the speaker-independent recognition technique cannot requires training by the users and easy to use. With the help of a modified real time voice activity detection algorithm （VAD） based on the log-energy acceleration associated with voice onset, we only perform feature extraction and decoding to the active voice and ignore the frames of non-activity. To further decrease the computational loads, we analyze and optimize to the calculation of state output probabilities. Test on 100 words vocabulary shows that system provides a recognition accuracy rate of 98.1% using only 1.03 times of real time.