Research On Anti-Noise Of Speech Recognition Based On Continuous Hidden Markov Model

With the advent of the information age,Hidden Markov Model(HMM)is widely used in the field of speech recognition,and has become one of the most concerned and effective modeling methods in this field.As people have higher and higher demands on intelligence,HMM-based speech recognition technology is also constantly improving.Especially in noisy environments,the original recognition system and the real speech cannot be highly matched,which greatly reduces the system recognition rate.Therefore,this paper will study the reliability and anti-noise ability of speech recognition system based on HMM continuous situation(CHMM).This paper first improves the CHMM based on the maximum mutual information(MMI)training criterion,and initially improves the recognition rate of the system.Then in the background of white noise in signal space research,a linear prediction HMM(LPHMM)is proposed,which uses Gaussian colored noise whitening method(GCNW)and minimum mean square error estimation(MMSE)to obtain a speech enhancement method(LWM),which is compared with the classical speech enhancement algorithm spectral subtraction(SS);in the feature space,based on the feature parameter Mel frequency cepstrum coefficient(MFCC),it is optimized by weighting,combined with the time-domain feature short-term energy(En)and short-term average zero-crossing rate(ZCR)to obtain a feature parameter Extraction method(MFFEZ);in the model space,based on the model compensation method(MACA)for joint compensation of additive noise and channel function,combined with MFFEZ feature extraction,the MFFEZ-MCAC model is obtained,and its identification with the MACA model in a noisy environment is compared rate.Finally,Matlab is used for experimental simulation.The research shows that when white noise is mixed in the speech signal,the LWM method has better anti-noise effect than spectral subtraction.The MFFEZ-MCAC model has a higher recognition rate than the MCAC model in mixed noise speech recognition.