Research Of Speaker Verification In The Channel Mismatch Conditions

With the development of science and technology, the research of speaker verification technology is refocused from laboratory environment to the complex environment in real life. It brings about many new problems to be solved. Channel mismatch is one of the typical and representative problems. Channel mismatch that training voices and testing voices from different transmission channels results in the decline of the system of speaker verification.The thesis analyzes two speaker verification systems. One is based on GMM-UBM structure, and the other is based on SVM. It discusses common channel mismatch compensation algorithm of feature, model and score. The solution of the problem is proposed, which is based on nuisance attribute projection (NAP). Then the algorithm is optimized.The main research contents are as follows:Firstly, the thesis discuss the speaker verification system based on probability statistical GMM-UBM. Then EM algorithm and MAP algorithm are discussed deeply. Aim at improving the distinction of probability statistical model system and disposing the problem that the lack of speaker personalized information description of distinguish identify model, the GMM-SVM speaker verification system is proposed. GMM model is used to compress and cluster the feature parameters. GMM-supervector from GMM is used as the input of SVM to set up the speaker model.Secondly, the channel compensation mismatch method of speaker verification is discussed, such as CMS, RASTA, Feature Mapping, Factor Analysis and T-norm. Comparative experiment based on NIST database shows that the methods which are discussed above can improve the negative impacts of speaker verification which is brought about by channel mismatch.Thirdly, the problem of channel mismatch from GMM-SVM speaker verification system is explored deeply. The mismatch compensation algorithm is proposed that it can remove the information of channel of GMM-supervector used as SVM input. Voices from a number of known channels are used to train projected matrix. Then training voices and testing voices are projected using the matrix. Last, the speaker verification system which is suffering little effect from channel appears.