The Research Of Speaker Recognition Base On GMM

Speaker Recognition (SR) is one of Biometric authentication technology, Its purpose is through the analysis of voice information to identify or confirm the identity of the speakers. This technology has the advantages of safety, fast and economic. SR has a great value in many fields such as electronic commerce, banking, justice, and military.The speaker model is obtained by the training of the voice, then use the phonetic characteristics under test to match with the models, finally judge the ID of the speaker. Thus, in order to achieve speaker recognition need to solve the following basic questions:(1)the preprocessing of speech signal; (2) training speaker model;(3) the calculation of model matching. In-depth study on several aspects in this paper, achieve the goal of improving system performance. Traditional speech endpoint detection method is useless in complex environments, so this paper presented a endpoint detection method base on spectrum analysis. From language spectra, distinguish between voice and noise by divide threshold. Combine the Multiple distinguishing of wavelet transform and the time-frequency analysis characteristics of spectrogram, the paper presented a de-noising method based on wavelet-spectrum analysis. As the number of the model increasing, recognition rate will decrease gradually, results in a real-time system is abate, and can’t meet the requirements of users. Papers in hierarchical identification is proposed based on the model under the thought of the density of the speaker clustering method, clustering to solve the following two questions:(1) Unsupervised clustering often make the result is not reasonable, and can shorten the recognition time.(2)The speaker model is more parameters mode, how to effectively measure the degree of similarity between two models is very important. For the problem one, the paper put forward the strategies of the evenly divided, so that when the clustering of membership is in the set range. In view of the problem two, this paper puts forward the concept of density of model, and combining the approximate KL divergence degree of similarity between common measurement models.The experimental results show that the proposed wavelet spectra de-noising method can adapt to a variety of language environment, meet the requirements of speaker recognition. The proposed clustering method based on the model of density model, the model can ensure the recognition rate loss is less than 1% of cases, recognition time shortened to a quarter of the original. So, This study play an important role for improving the real-time performance of the speaker technology.In addition, the paper also introduces the speech front-end processing, the principle of cepstrum characteristic parameters and Implementation steps, the limitations resolution strategy of Gaussian mixture model, etc.