Research On Speaker Recognition Based On Fusion Feature And Gaussian Mixture Model

With the rapid development of mobile Internet,convenient and efficient biometric authentication technology has been paid more and more attention.Speakers recognition also known as voiceprint recognition,is a kind of biometric authentication technology,widely used in communications networks,commercial transactions,forensic and military Security and other fields which by virtue of more convenient,safer,more economical advantage ushered in the development of the opportunity.The key of speaker recognition is feature extraction and model training,the commonly used speech features is linear predictive cepstrum coefficients(LPCC),Mel frequency cepstrum coefficients(MFCC)and perceptual linear predictive(PLP),which the recognition performance of MFCC and PLP is better than that of LPCC.However,MFCC and PLP are used as a single feature to identify the speaker's personality and still can not meet the high accuracy requirements of some fields.The Gauss mixture model(GMM)is the most widely used because of its simplicity and good recognition performance.But the simple initial parameter selection is simple and easy to identify method to make it unstable,lack of accuracy and other defects.Based on this paper,a new fusion feature PLP-MFCC and improved GMM are proposed to improve the performance of the recognition.The main contents are as follows:Firstly,the characteristic parameters LPCC,MFCC,PLP and their extraction process commonly used in speaker recognition are introduced.After analyzing the F and D ratio evaluation methods,The new fusion feature PLP-MFCC is obtained by fusing some of the characteristics of MFCC and PLP by F ratio and D ratio.Secondly,the modeling method of GMM is studied,the process of GMM training model is the process of parameter estimation.The selection of the initial parameters in the parameter estimation will affect the accuracy of the training model.The initial parameters are usually selected by random or K-means.After analyzing some of the problems,An improved K-means algorithm is applied to GMM parameter estimation to improve GMM.Finally,the speakers recognition system based on the Gaussian mixture model is constructed on the Matlab platform.Using the standard TIMIT speech database for experiments,The results show that the recognition rate of PLP-MFCC compared to MFCC,PLP were increased by 4.9%,3.8%.And then compared the GMM before and after the improvement of the modified GMM compared to the improved GMM recognition rate increased by 2.3%.The experimental results show that the fusion feature proposed in this paper and the method of improving GMM are a certain effective for improving the speakers recognition system.