基于DHMM的语音识别算法及DSP实现

Speech Recognition Algorithm Based on DHMM and Its DSP Implementation

针对非特定人孤立词的快速高精度语音识别问题,在分析语音信号特征和离散隐形马尔科夫模型(DHMM)原理的基础上,提出了一种基于DHMM快速语音识别算法。算法首先对语音信号进行预处理和端点检测以提取有用信号,进而完成语音特征参数提取和矢量量化,最后采用DHMM模型利用Viterbi搜索算法实现语音识别;在训练阶段需要完成码本设计和DHMM模型参数的生成;采用ADI公司Blackfin系列BF533芯片对识别算法进行DSP实现。理论分析和仿真结果表明,在日常环境条件下算法的识别正确率达97.5%,且DSP运行时间小于10 ms。

The paper focuses on the fast and high-accuracy speech recognition of speaker-independent isolated words. Based on an analysis of the characteristics of speech signal and the principles of Discrete Hidden Markov Model（ DHMM）,a fast speech recognition algorithm using DHMM is presented. Firstly,the useful speech signals are separated from the background noise by using preprocessing and endpoint detection. Secondly,speech feature parameters are extracted and vector quantized. Finally,speech recognition is realized by using Viterbi search algorithm based on DHMM parameters. During the training phase,the codebook design and generation of DHMM parameters are completed. The algorithm is DSP implemented by using the BF533 chip of Analog Devices Inc. Theoretical analysis and simulations show that the recognition accuracy is as high as 97. 5% under daily environment conditions and its DSP runtime is less than10 ms.