Rearch On Text-independent Speaker Identification Technology Based On SVM

Speech,the basic basis for speaker recognition technology because of its uniqueness,is the most effective way for human to communicate with each other.Under the framework of speaker recognition,it is the research of hot spots in the field of speaker recognition to find a kind of discrimination feature to obtain higher system performance.Model selection and feature extraction are the key issues in speaker recognition technology,and the performance of speaker recognition system depends on the type of feature under the determining model.In the today's digital era,it is of great theoretical and practical significance to find a superior speaker feature.The goal of this thesis is to design a speech feature that enables the speaker recognition system obtain the better recognition performance and the lower time complexity.The recombination Supervector is proposed in the thesis after researching the characteristics of GMM Supervector in speaker recognition system,and the feasibility of recombination Supervector is analyzed depends on the characteristics of SVM.Then a deep neural network is designed to extract the bottleneck feature of speaker speech after studying the deep learning.The main research content and innovation can be showed in the following.(1)The basic theory framework of speaker recognition is introduced in the thesis,including speech signal preprocessing method,feature extraction method and speaker recognition model.Not only the extraction process of LPC and MFCC and their cepstrum features is introduced detailed,but also the analyses of their characteristics.In addition,the classical speaker recognition methods such as template matching algorithm,HMM method,VQ method,GMM method,SVM method and DNN method are also introduced.Through the previous research,there is a discovery that the performance of the latter three methods are relatively better in the speaker recognition system,so the research of this thesis is based on them.(2)Inder to overcome the problem that the performance of traditonal Supervector is not well enough,the GMM-SVM text-independent speaker identification system based on recombination Supervector was proposed in this thesis.And the recombination Supervector can represent the inner detail of speakers' identity better,which enable the new system take the advantage of the characteristics of superior performance when SVM deal with the small and high dimensional data.The experimental results demonstrate that the GMM-SVM speaker identification system based on recombination supervector can not only obtain a higher recognition rate than the traditional speaker identification system based on GMM-SVM,but also shorten the implementation time greatly.(3)To deal with the problem that the depth structure information of the speaker speech can not digged by the traditional feature,this thesis designs a depth neural network to extract the bottleneck feature of the speaker speech,and establish the speaker identification system based on DNN-SVM.The bottleneck feature that with invariance and superior distinction can dig the depth feature of speaker.The experimental result demonstrate that the speaker identification system based on DNN-SVM obtain a better recognition rate performance than the speaker recognition system based on SVM.