Speaker Recognition Research Based On Clustering Analysis And Neural Network Ensemble

Speaker Recognition technology which is based on the characteristics of thespeech signal to automatically recognized the speaker’s identity.As a biometricauthentication technology, compared with other technologies which is moreconvenient, economic, and safe. It is widely used in the network, defense, security andother fields. Therefore, the study on speaker recognition technology have universalvalue and important practical significance.At the beginning of this article, the Speaker Recognition is to be overview. Gavean outline of its basic principles, recognition system structure and implementationprocess. First, we analysised the two mainstream characteristic parameters of speakerrecognition system, LPCC and MFCC. Secondly, the knowledge of the theoreticalbackground to be used the content of this paper-cluster analysis and neural networkensemble were full and detailed description. Next, for the defects of the traditionalk-means vector clustering algorithm for speech signal proposed based on the varianceweighted geometric distance to improve the traditional k-means algorithm, eachcomponent of the feature vector weighted by variance size, the weighting factor is thereciprocal of vector variance. After the improved algorithm combined withBP_Adaboost integration model, applied to both speaker identification systemcombination. Both are applied to speaker recognition system.Lastly we made the recognition system of this paper used in Matlab software forthe simulation experiment. First, the speech signal were pre-processing and the other aseries of processing were done. Because the Mel cepstrum in reflecting the human earon the auditory characteristics of an edge is more advantage, so, in terms of choice ofthe characteristic parameters, we chose Mel cepstrum coefficients as parameters.Made the MFCC as the input vector of the training and identification of the systemmodel to train and learn. Finally, made the speaker recognition system based on themethod of this paper, compared with the Speaker Recognition that based on the BPneural network model and that based on the BP_Adaboost integrated model in orderto verify the validity of the method in this article.