Research On Channel Algorithm Of Speaker Recognition Based On GMM-SVM

The direct way to communicate with each other among people is the voice, and the voicesignal disseminates useful information, carries the information, and has a multi-level researchvalue. The voice is a direct communication tool, and the characteristics of different humanspeech organs and other differences caused by such as the environment, making it animportant tool for identification. Speaker’s voice signal carries specific different characteristicinformation, conveys the speaker’s semantic content, and also contains the specificphysiological characteristics, mood and other important personal information, making it highvalue for research and practical applications. As more and more researchers are dedicated tospeech recognition, its application will be more valuable.Speaker recognition is a kind of biometric identification, which refers to extracting thespeaker’s personality characteristics from the speaker’s speech, exceeding the analysis andidentification through the personality characteristics, so as to achieve the purpose ofidentifying or confirming the speaker. Speaker recognition is considered to be the mostnatural way of biometrics identification. Because the voice is the inherent characteristic of aperson, can be generated naturally, and the process of training and recognition does not needspecial input devices, you will easily find the devices that can be used as an input device suchas the microphone of personal computers and mobile-phones, so the price of speakerrecognition system is low. These reasons make voice a popular biological characteristic.Although the speaker recognition technology has made considerable progress, but theapplications of speaker recognition actually has a lot of problems to be solved.With the continuous development of speech recognition technology, researchers haveproposed a variety of speech feature parameters. And dynamic time warping algorithm, vectorquantization, Hidden Markov chain are also introduced into speaker recognition. The papercombines the Gaussian Mixture Model and Support Vector Machine for speaker recognition,which has better description for speakers and strong distinction performance. Compared withthe previous algorithm, the GMM-SVM speaker recognition has better ability to characterizedata, better classification performance, and better recognition efficiency.Further research let me find that, the recognition rate to be improved is effected by thechannel factors seriously. To compensate for the shortcoming of GMM-SVM speaker recognition, the article studies the feature level and model level channel compensationtechniques:(1) For the problem of that the speaker speech feature is susceptible to additive noise andlinear channel interference, which causes the mismatch of channel, this paper proposes the useof features warp bending on voice data. Apply compensation on features level such as theFeature Warping into the GMM-SVM speaker recognition in the paper to make frequencycepstrum features more pure and robust. Features warping take use of cumulative distributionfunction to make the feature vector sequence characterized to the standard distribution,enhance the channel robustness of voice features and anti-noise performance, and can help toget a better robustness in the gradient noise environment and channel mismatch situations.Features Warp is applied to speaker recognition as a normalization method, improving therobustness from channel of speaker recognition, strengthening its performance to adapt todifferent environment, so as to improve the performance and accuracy of speaker recognition.(2) Speaker model is established with the speaker feature vectors which are associatedwith the channel, so inevitably the model is interfering by channel information components,then the accuracy of speaker recognition is effected. To solve this problem, this paperproposes the use of factor analysis algorithm. The algorithm separate the speaker speakermodel into the speaker space and the channel space. With the method, eliminate redundantcomponents in cep-strum features. Divide the hidden channel information beneath the surfaceof the speaker data, then obtain the speaker-corresponding space and the feature vector.Ultimately take speaker model irrelevant to channel as SVM training data. The experimentalresults show that the application of the model compensation based on factor analysistechnology guarantee higher recognition efficiency and better channel robustness than thetraditional GMM-SVM speaker recognition.The paper makes the detailed introduction of the design framework, principles ofalgorithms and so on, which is based on the basic principle of GMM-SVM speakerrecognition and channel compensation algorithm on feature level and model level.Experimental results show that the algorithm performance has been improved. Featurewarp effectively reduces channel interference cepstral features and the impact of signaldistortion for speaker recognition, thereby enhancing the recognition effect. Factor analysisestimate channel factor of the Gaussian super-vector and the eliminate it, thereby reducingchannel influence and improving the performance of the algorithm.