Study On The Key Technology Of The Speech Recognition And It’s Improved Algorithm

As one of the important fields in the research of computer technology, speech recognitionhas great research value and widely application prospects. But most speech recognitionsystems at present are still limited in lab. Although there have been some products of thespeech recognition at the market，there is also a large gap between the people’s expectationand the actual use effect.This thesis focus on the two keys of the speech recognition system: efficiency andaccuracy. Along with the flow of the technology of speech recognition，this thesis goes deepinto the research on the key points of speech recognition．Firstly, the thesis introduces the basic principle and process of the speech recognition.Then the thesis analysis the digital model of speech signal and the problems of itspreprocessing. For the bugs of the traditional voice end point detection, a new method of thevoice end point detection based on the image edge detection is proposed in this thesis. Theresults of the experiment show that the new algorithm improves the detection accuracy rate by18.6%. It is superior to the traditional algorithm．Secondly, the characteristics and extraction process of several common speech featureparameters are introduced and discussed in the thesis. Based on the in-depth study of MFCCparameter, this thesis propose an improved MFCC parameter after a series of processing onMFCC feature such as weighting, differential, and PCA dimensionality reduction. Theexperiment results demonstrate that the new feature parameter makes the speech recognitionsystem more robust. The average recognition accuracy in three kinds of SNR environmentsincreases9.2%and4.3%respectively compared to the traditional feature parameters thatLPCC and MFCC. The average training time has been shortened by18.2%and11.5%respectively.This thesis researches on the technology of speech recognition based on HMMemphatically. Owing to the application of the HMM in the research of speech recognition,the speech recognition has made considerable headway．The HMM has already been the mostwidely used modeling technology in speech recognition currently. For the traditional HMMmodel initialization is too simple and crude, the thesis proposed a new improved HMM model initialization algorithm. The results of the experiment prove that the new model initializationalgorithm can reduce average training time by36.9%and improve the system recognition rateby5.2%.Finally, on the basis of the theoretical research on speech recognition, using MATLABsimulation software and VoiceBox speech signal processing toolbox, the thesis establishes asmall speech recognition simulation system which is used to do the experiment. For isolateword speech recognition, the recognition accuracy rate of the system has reached95.5%.