Research On Text-Independent Speaker Verification System

Automatic speaker recognition is the processing of automatically recognizing who is speaking by using speaker specific information included in speech signal. With the development of communication and information technology, it is getting more and more attention for its bright future. It can be classified into speaker identification and speaker verification according to decision modes. This thesis focuses attention on text-independent speaker verification. The main works are as follows:1) We have a deep research in basic theories of speaker recognition and the speaker verification system based on Gaussian Mixture Model.2) In order to improve the performance of Gaussian Mixture Model, we use hypothesize test to do the speaker verification. We introduce two anti-models: Universal Background Model and Cohort Model. Considering the feather of the two models, we combine them using a new combination method.3) We try to import the Bayesian adaptation, which is widely used in speech recognition, into speaker verification. We use Bayesian maximum a posteriori estimation training a speaker model from background model, to solve the problem of model miss matching in speaker verification system.4) We setup a text-independent speech corpus, which include 50 people. Basing on this corpus, we test both the two improved method and give the detail test results. And the results show the above two methods have a better recognition performance than the origin Gaussian Mixture Model. There are little difference in the recognition performance between the two model.