Research On The Speaker Recognition Based On VQ And GMM

Starting in 1930's, speaker recognition has increasingly become a hotspot of research since 1960's. It can be applied to a number of fields, such as security, justice, military affairs, finance and services. Because of that, lots of scientific researchers are involved in the research, making great development. However it is not ripe very much.This paper is mainly about a text_dependent speaker recognition system based on vector quantification (VQ) methods. We use LPC-derived cepstral coefficients, delta-cepstral coefficients, pitches and delta-pitches as the feature parameter set. Through the test of a speech library composed of 10 speakers and 1800 speeches, we get the fairly high recognition rate, which is about 92% for the one-syllable words, about 96.67% for the two-syllable words and about 97.67% for the four-syllable words. This system can also be used in the real-time speaker recognition. The experiment result reveals to be with real-time, satisfied recognition rate (above 90%).when Gaussian mixture model(GMM) is the best model of text-independent speaker recognition, the change of environment and thumbprint have an effect on the recognize ratio of GMM which adopt standard maximum likehood. In this paper, according to the statistic characteristic of frame likehoods, we describe a non-linear transform method- exponential normalization transformation as a technique used in speaker recognition. The theory and experiment show that the exponential normalization transformation could improve the recognize ratio.This paper consists of five parts. The first and second parts mainly discuss the common parameters of phonetic features and extraction methods. The third and forth parts focus on speaker recognition methods, based on VQ and GMM. In the last part, we discuss the realization of the speaker recognition system and the results of the speaker recognition experiments.