Telephone Channel, Text-independent Speaker Recognition,

Speaker recognition is the process of automatically recognizing who is speaking on the basis of individual information included in speech signals. The speaker identification technique can be well applied in many fields with bright future. Based on the overviews of the history of the development on the speaker recognition technology and the explanation of the basic principles of the speaker identification, this paper put special emphasis on the feature extraction, the modal selection and the decision rule in speaker recognition technique. In transmission system, the channel noise can not be ignored. The paper proposes the feature mapping to build the relationship between the features extracted from different channels.The core of reorganization problem is classifier design. The paper presents text-independent classifiers: GMM classifier and VQ classifier and proposes the dynamic recognition length choose algorithm (DRLCA) based on optimal stopping rules in detail.The basic structure of this recognition system is multi-classifier cooperation. There are many methods that can make classifiers cooperate. The paper presents the linear combination method and competition method, and proposes the evidence fusion method. The traditional D-S is not fit for speaker recognition application; The paper proposes close-set fusion and open-set fusion functions to meet the different kinds of speaker recognition application requirement.At last, the paper proposes a applied system based the techniques above. It can supervise transmission channels and give out recognition results. The system works on the condition of Odb and 13 speakers. The recognition ratio is 94%.The system faces to application, and tries to overcome the noise affection and shorten the recognition time. My main work is as following: 1) applying feature mapping, sub-band structure classifier and multi-classifier cooperation to enhance the robust of system; 2) giving out close-set fusion and open-set fusion functions to solve the problems of speaker identification and verification respectively; 3) building the dynamic recognition length algorithm based on optimal stopping rules; 4) developing a applied system based on the techniques above.