Time Varying Vector Auto-regressive Modeling With Application To Speaker Recognition

Speaker recognition is a technology identifying the speaker automatically accord-ing to the voice parameters which imply the speaker’s physiological and behavioral characteristics in the speech signal. It is a secure and stable bio-authentication tech-nology and has broad application prospects. How to extract the speaker’s personality traits accurately from the speech signal is the key to the problem of speaker recogni-tion.This paper focuses on text-dependent speaker recognition. By preprocessing the speech signal, analyzing in frequency domain and cepstrum domain, the average MEL cepstrum is got. The frequency corresponding to the MEL cepstrum peak is selected to be the characteristic frequency. Frequencies corresponding to the maximum time-varying and maximum MEL cepstrum peak are used as the characteristic frequencies in this study. The multiple linear regression equation is established to separate the deter-ministic components and stochastic components of the two characteristic frequencies Mel cepstrum value series.On the basis of TVPAR model theory, the time-varying vector auto-regressive (TVVAR) model is established to further extract the speaker’s speech signal parame-ters. The D ratio is selected to evaluate and select the final identification of parameters. The Mahalanobis distance is selected to recognize the speaker preliminarily and the ex-perimental recognition rate is99.8%. In order to further make use of the information, the distribution of the distance difference is analyzed, which is obtained according to the difference between the distances to the analyzed speaker’s template and to the other. By use of probability theory, the recognition rate in this experiment reaches to100%. And a reasonable credibility is given for every recognition.