改进的小波变换HMM语音识别算法

Speech recognition algorithm based on wavelet transform and improved HMM

语音识别系统的识别率十分依赖基于Hidden Markov Models(HMM)模型的训练技术.然而,经典的训练算法(Baum-Welch算法)有一个致命的缺陷,即所得最终解依赖于初始值的选取,只得局部最优解,这就影响了系统的最终识别率.针对传统语音识别系统识别率较低的现状,提出了一种改进的小波变换HMM语音识别算法.该算法首先通过小波变换对原始语音信号进行了降噪处理,然后使用语音样本对利用遗传算法改进后的HMM模型进行训练,并用于语音识别.实验结果表明:所提出的算法实用有效,识别率显著提高.

Recognition rate of speech recognition systems relied heavily on technology-based Hidden Markov Models-HMM model training.However the classic Baum-Welch training algorithm had a fatal flaw,namely,final solution obtained depended on the selection of the initial value,which was often only locally optimized solution.It would affect the recognition rate of the final system.To increase the recognition rate of traditional speech recognition system,it was presented an improved algorithm based on wavelet transform and HMM model.Firstly,noise in the original signal was reduced by wavelet transform,then an improved HMM model trained by speech samples and used to recognize speech.Experimental results showed that the improved algorithm,which was implemented by genetic algorithm,was practical,effective and system recognition rate was increased significantly.