Research On Speaker Recognition Based On Combination Of Features

Speaker Recognition is a kind of biometrics technology depending on feature parameters extracted from his/her speech to identify speaker. It has quite a wide application in many security areas such as information security, communication, justice and military for its particular advantage on convenience, economy and veracity. Feature extraction and pattern matching were the two key of speaker recognition, and the problem was mainly to extract speaker characteristic parameter based on a large number of researches. So this thesis mainly focused on dealing with the speaker feature extraction on the base of analyzing and studying the basic principles of speaker recognition system. The main work were as follows:(1)Although the speaker recognition ratio was gotten some increase by combination of Linear Prediction Coefficient(LPC) and Mel Frequency Cepstrum Coefficient(MFCC), but it will increase the dimensions of the feature parameters and bring about heavy computation. Aim at this problem, the method by integrating LPC Coefficient into the computation of MFCC Coefficient was presented in this paper. Firstly, LPC parameters from speech signal were calculated and the speech power spectrum of LPC were gotten; Secondly, the logarithm of output by making the speech power spectrum of LPC through triangular filter group were conducted. Finally, the output of logarithm was transformed by discrete cosine transform, and a new feature factor which is called Linear Prediction Mel Frequency Cepstrum Coefficient(LPMFCC) was obtained. LPMFCC parameters had both vocal track of LPC parameters and auditory of MFCC parameters. Although increasing a step of computation, the dimension of parameters was not increased and computation costs was relatively low. Simulation experiment was made by applying the LPMFCC method to speaker recognition system based on GMM and VQ. In the pure voice database, the experiment results show that the recognition rate of the proposed LPMFCC method rises by 18.57% and 10% than the recognition rate of LPC method and MFCC method respectively for the speaker recognition system based VQ; and rises by 11.72% and 2% respectively for the speaker recognition system based GMM. Furthermore, the proposed LPMFCC method can obviously improve recognition performance in various noise environments.(2)Mel filter banks were dense in low frequency area, while it were sparse in high frequency area, so the high-frequency information was ignored by MFCC coefficient. In view of this insufficiency, the method inverting the Mel filter bank structure was used, and Invert Mel Frequency Cepstrum Coefficient(IMFCC) was obtained. Considering the complementary relationship of the MFCC coefficient and IMFCC coefficient, application of integrating feature from MFCC and IMFCC in system was discussed. Firstly features of MFCC and IMFCC were input into same classifier, and their own match scores were gotten. Secondly the fusion match score was obtained by adopting weighted fusion their score. Finally verdicts were given. the efficiency of the method by experimental simulation was verified.(3) The problem in the application of speaker recognition technique was researched. In this paper, a DSP-based speaker verification system was initially realized, which identity the speaker by LED in the development board, and the light is on when the speech is from oneself, if the light is not on, it means the speech is not from oneself.