Research On Speaker-Independent Speech Recognition System Based On HMM

As one of the most difficult speech recognition systems,speaker-independent speech recognition has been under intense discussion and research exploration.However,how to ensure the high recognition rate of the speech recognition system is always a hot topic in scientific research under the extremely uncertain conditions of the speaker and the pronunciation environment.This thesis studies the constituent elements of speaker-independent speech recognition system,and mainly puts forward an improvement scheme for speech enhancement algorithm to improve the stability of speech recognition system in noisy environment.Finally,the system is built on MATLAB platform.The main key points of this thesis are as follows:1.This thesis first focuses on analyzing several speech enhancement algorithms,and finally combines wavelet threshold denoising algorithm with Kalman filter,which improves the threshold function.The experimental analysis shows that the method has good filtering and noise reduction effect on noisy speech signals mixed with different background noises.2.In the pre-processing of speech signals,two improved speech endpoint detection algorithms are given respectively by using different time domain and frequency domain parameter characteristics.Through experimental analysis,the time domain parameter method based on short-term average energy(amp)and short-term average zero-crossing rate(zcr)is better for continuous word recognition.However.the frequency domain parameter method based on MFCC cepstrum distance can identify word for word better.Finally,combined with the actual situation of this system,the improved time domain parameter method is selected as the endpoint detection algorithm of this identification system.3.Two commonly used speech feature parameters,linear predictive cepstrum coefficient(LPCC)and Mel frequency cepstrum coefficient(MFCC),are analyzed The advantages and disadvantages of the two methods are compared,and the MFCC extraction process is introduced in detail.4.In the aspect of speech recognition module,this thesis gives an improved way to initialize HMM model parameter B.The experimental analysis shows that the method effectively reduces the iteration times of speech training.Finally,HMM model is combined with the improved speech processing module to build a high recognition rate speech recognition system that can work in different background noise environments.Finally,in order to test the recognition effect of the speech recognition system established in this thesis,we have built a small speech library containing 2250 voices by using Cool Edit Pro tool.After experimental analysis,under the condition of pure speech,the recognition rate of the speaker-independent speech recognition system based on HMM established in this article is above 98%.However,high recognition rate of more than 93%has been achieved in the low SNR environment mixed with noise.