The Study Of Advanced Kernel Function Algorithm And The Application In Speaker Identification

Due to its special merits of flexibility, economy and accuracy, speaker recognition technology has a broad application future in biometrics identity verification field. However, speaker recognition has some limits in application because the training algorithm is complicated, and the robustness is not ideal. Support Vector Machine (SVM) is a new classification methodology. It has been proved to be a powerful technique in pattern classification for its good generalization ability. But SVM has some disadvantages in some aspect for it's still in the developing stage.The thesis focus on how to improve the recognition ratio and robustness of speaker recognition system by generating new kernels based on supervector. The main contributions of the dissertation are as follows:(1) The advanced feature parameter extraction. This thesis introduces Gaussian Mixture Universal Background Model (GMM-UBM) into speaker recognition modeling. UBM uses speaker-independent distribution parameters to approximate parameters for acoustic units which are absent in specified speaker's training data. Then stack the means of the GMM model which is adapted by MAP algorithm to form GMM mean supervector.(2) The adoption of new kernels, such as Kullback Leibler Divergence kernel, L~2 inner product kernel and NAP kernel. These three new kernels are all based on GMM supervector. The SVM using the kernels based on GMM supervector can be used to classify wholly on the sequence. Meanwhile, in order to enhance the robustness of the system, the thesis adopt kernel based on Nuisance Attribute Projection. This kind of kernel gets rid of redundant information from the subspace which has no relationship with the speaker feature.(3) Based on people voice database, we take emulation experiments. First, we compare the advanced feature extraction method with RBF kernel and polynomial kernel. Second, we apply these three new kernels to speaker identification. From the result, we can see these three new kernels improved recognition ratio at least by 12%, and NAP kernel improved the robustness of the system a lot.