The Study Of Speaker Verification Based Channel Compensation Methods Between Training And Testing Data

With full speed development of information today, many occasions require to affirm and recognize the speaker's identity. In recent years, the emphasis of speaker verification has in a minute changes to the actual complicated environment gradually from the laboratory, the problem caused by the different channels between training data and testing data was emphasized by people, and became the key point of the speaker verification research. In telephone network, the speech of speaker not only come from the different channels, but also come from different phone or microphone. In the system of speaker verification, the system carries out an erroneous judgement when the speaker model is trained by certain channel while testing speech comes from other channel, so that it reduces recognition rate of the speaker verification.The text introduces some common methods of compensation of channel mismatch, including feature field,model field,test-score field; then deeply discusses the usage of SVM for speaker verification system; last for better decision-making in a text-independent speaker verification system, performance degradation caused by mismatch that occurs between training and testing, the channel compensation by high-dimensional projection was introduced and a new projected support vector machines model was proposed through the based system, the context of main research as follow:First, the text researches the speaker verification system based on probability statistical GMM-UBM, deeply understands the arithmetic of GMM training and MAP, analyzes the problem of channel mismatch of GMM-UBM. The text deeply explores the method of compensation of channel mismatch in the complicated circumstances as CMS,RASTA,Feature Mapping,Z-norm,T-norm and Factor Analysis which is the focus of research at present.Second, The text explores the method that constitutes SVM model by adopting GMM super vector, Support Vector Machine develops from the most super plane which is linear apart optimization. The speaker verification is a problem of two classes, target speaker as +1 classier, imposter speaker as -1 classier in order to establish SVM model. In this way, it can reduce the quantity of compute and wipe off the redundant information of feature parameter, improve the performance of the speaker verification effectively.Third, the text analyzes the problem of SVM training and testing caused by the different channels, proposes a new method of compensation mismatch based on high-dimension projection and gives a new speaker verification system of channel compensation of Projected SVM. GMM super vector is compensated by a projected matrix, then project it to a high-dimension space by PSVM kernel function in order to constitute PSVM model which is less affected by channels. The projected matrix is high-dimension matrix, it is trained through distance to describe different channels'difference by many other speech which have known about the channel, then can acquire the channel space and speaker space.The text proposes a new PSVM model instead of SVM model, the method basically reduces the affect of channel information for speaker verification, and improves the recognition rate of the system.