Adaptive Gaussian Mixture Model And Its Application In Speaker Recognition

Language as the most convenient and effective tool in exchanging information withexternal worlds, plays an increasingly important position in daily communication. Due tothe complex speaker characteristic features, traditional researches did not achieveleapfrog development in a long time. While there are still some institutions and scholarsraising lots of merged and improved methods to further related studies, to promote thedevelopment of speaker recognition technologies. However, the under-fitting andover-fitting problems that often occur in model fitting, have not been satisfactorilyresolved. The accuracy of model fitting still needs further investigation.This paper establishes a speaker recognition system based on Gaussian mixturemodel to evaluate the speech recognition performance in the different length of clean testenvironment with different parameters of LPCC, MFCC and BFCC, proving that BFCCcompared with MFCC and LPCC has a lowest error rate. An iterative self-organizingGaussian Mixture Model (ISO-GMM) and adaptive Gaussian Mixture Model algorithm(AGMM) are proposed to optimize the fitting problems of traditional GMM. ISO-GMM,a method using the cluster characteristics of ISODATA, establishes the appropriate modelin the training process according to the specific distribution of speech signals.Experimental results show that the ISO-GMM method is more in line with the diversityof voice acoustic feature distribution to help improve the recognition rate than traditionalGMM. The relative error rates of MFCC and BFCC decline by19.79%and11.47%respectively. The mixture number of AGMM is variable in speak recognition. And withthe cluster property of acoustic feature distribution, it adopts an absorb-merge-splitmechanism to adjust the mixture numbers dynamically in the model training to achievebetter model fitting performance. Comparisons of the recognition performance betweenthe traditional GMM and AGMM in different test speech length show that the fitting performance of AGMM is more superior than the traditional GMM, which solves theover-fitting and under fitting problem. The relative error rates of MFCC and BFCCdecline by41.41%and22.21%respectively. In the end of the paper, a completedcomparison between ISO-GMM and AGMM is presented.