Analysis And Research On Speaker Recognition Based On Vector Quantization

Speaker recognition belongs to biometrics technology. It is the technique to recognize the speaker identity automatically on the basis of the individual's physiological and behavioral characteristics included in speech waves. It has a broad application prospect in many fields. This thesis focuses attention on speaker recognition based on vector quantization.Firstly, feature parameters of the speech signal are analyzed in this thesis, and then the elementary theories and extraction methods of LPC,LPCC and MFCC are discussed in detail. Secondly, this thesis mainly studies the speaker recognition based on VQ. According to the LBG algorithm, presents a new arithmetic for initial codebook design; according to the matching decision, presents a weighted Euclid's distance distortion measure based on the standard deviation.During the experiments, it is shown that weighted MFCC is a kind of well-performance feature parameter, which could ensure high recognition ratio for the system; experiments have validated the correctness and reliability of the codebook design arithmetic, and reviewed the influence of codebook dimension on the system recognition performance; experiments have also tested two different weighted strategies, one is weighted feature parameter, and the other is weighted recognition method, which both have improved the system performance. Lastly, a test platform of text-independent speaker identification system is established under the Matlab software. Experiments have tested on a speech database contained ten speakers, obtained a much higher recognition ratio.