Speaker Verification Based On Sorted GMM

With the continuous progress of society and civilization, the way of communication among people are more diversity, to identify the specific identification of the speaker's identity are increasingly high demands. In all the technology of authentication, the biometric technology is very widely used because of this technology is based on physiology and behavior of human beings, this show biometric authentication technology has broad prospects for practical application. In the existing biometric authentication technology, the text-independent speaker verification is one of the most naturally biometric authentication technology, it identify the speaker's identity by the speech of the specified speaker. text-independent speaker verification is a very important research direction in speech recognition. Most text-independent speaker verification system use short-time cepstrum coefficients and the architecture of GMM-UBM-MAP. The text-independent speaker verification system which use this architecture has achieved high recognition performance already.The system's recognition performance and computation are two important criteria when we select speaker verification system. In the training process for UBM, for each input feature vector, we should compute all the likelihood value of Gaussian component in UBM, because the sum of Gaussian component in UBM is very big, the computation of training UBM is very large. To some extent, this limits the text-independent speaker verification system which used architecture of GMM-UBM in practical. The text discuss a text-independent speaker verification system which use speech of practical mobile phone or telephone, analyze the advantage of the training algorithm and architecture of GMM-UBM-MAP,then deeply research the method to reduce the computation and increase the speed for training UBM. The context of main research as follow:First, introduce speaker verification system which based on the architecture of GMM-UBM in detail, deeply discuss the arithmetic of GMM training and MAP. The system based on GMM-UBM-MAP has good offset influence to background of test speech, it show the information of the target speaker's personality, so the system based on GMM-UBM-MAP has better performance and noise robustness than the system based on GMM alone. Second, introduce the SVM model which it has good discriminative feature, and compare the text-independent speaker verification based on GMM-UBM-MAP and GMM-Sup-SVM.Third, introduce two extraction method about the track cepstrum parameters MFCC,LPCC based on short-time analysis, and discuss these two parameters about its effectiveness and robustness in speaker verification system. The experiment results show that these two parameters have good recognition performance, but compare with LPCC, MFCC has better performance.Fourth, to reduce the computation and increase the speed for training UBM, present a training UBM method based on sorted Gaussian mixture model, by sort the mixture components in UBM with pre-defined criteria, the input speech frame can only participate parts of mixture components, so it can reduce the computation of training UBM. The experiment results show that by using sorted Gaussian mixture model, we can reduce the computation, and almost has no influence on recognition performance.