The Research On Speaker Recognition In Noisy Environment

The speaker recognition rate is higher in pure voice environment,but decreases sharply in noisy environment. Now people propose various methods to improve the speaker recognition performance, but there are still a lot of problems to solve. The object of this paper is text-independent speaker recognition system, and study speaker recognition from multiple angles of noisy environment, research content mainly includes wavelet threshold denoising,combined feature extraction,two-stage decision model,preventing fake speaker invasion,and the speaker recognition system experiment results. This paper made the following several aspects of research:1.The wavelet analysis theory is studied and research wavelet threshold denoising algorithm and thought deeply, wavelet threshold function and threshold selection are studied,proposed an improved threshold function, the threshold adopts node threshold, noise variance estimation of traditional wavelet threshold value is based on white noise,in this paper the noise variance estimation of wavelet threshold value is calculated by spectral entropy method based on actual noise variance estimation, eliminating noise effectively and achieve speech enhancement purpose.2. The main research is feature extraction,this paper proposes a combined feature extraction method, extracting MFCC cepstrum characteristics in the feature extraction phase, forming combination characteristic which join the dynamic characteristics Delta Delta and the pitch based on average magnitude difference function, in the same experiment environment, contrasted with other characteristic features, the combined features is better than any other feature recognition.3.Model decision research, GMM is mainly applied to text-independent speaker recognition, DTW model is mainly used in text-dependent speaker recognition, so these two kinds of models complement and enchance each other, GMM and DTW model will form a two-stage decision model, in the stage of training, characteristic parameters are both trained by the two models respectively and stored model, in the stage of recognition, the characteristic parameters are identificated by GMM and output of 3 closest to the speaker's results, then through DTW model and choose one optimal result of 3 results, two-stage decision model can reduce recognition error rate. In order to prevent fake speakers invasion, joining score normalization befor judgment, excluding fake speakers before speaker recognition and improve the recognition rate.4. The speaker recognition system, including speaker recognition pretreatment, experimental speech database construction, the two experiments were made:experiment 1 is proved by Matlab that improved wavelet threshold denoising is superior to the traditional wavelet hard-threshold and soft-threshold denoising method;experiment 2 based on improved wavelet threshold denoising and two-stage judgment model of the speaker recognition method, pure voice recognition rate contrasted with improved wavelet threshold denoising speaker recognition, the method of the combined feature extraction and two-stage decision model, speaker recognition rate is contrasted which different feature characteristics in different recognition models and different features in different recognition model after adding score normalization,and analyzing the different experimental results.