The Study Of Support Vector Machine And Its Application On Speaker Recognition

Support Vector Machine (SVM) is a new and very promising classification technique. The approach is systematic and properly motivated by statistical learning theory. Training invovles separating the classes with a surface that maximizes the margin between them. An interesting property of this approach is that it is an approximate implementation of the Structural Risk Minimizaiton (SRM) induction principle.In this thesis, the theory and method of Support Vector Machines were studied in the given application, text-independent speaker recognition. Speaker recognition, which identifies or verifies people by their voice, is a kind of biometrics. Based on the recent advancements of Support Vector Machines and main points of speaker recognition, we proposed three related issues and discussed them respectively.The first issue is multi-class classification algorithm of Support Vector Machines. The traditional Support Vector Machines only deal with the binary classification. Unlike the one against one and the one against rest methods brought forward by other researches, we solved the multi-classification with two different approaches, one is constructing a uniform optimization function, the other is utilizing decision tree to combine the binary classifiers.The second issue is Support Vector Domain Description for open set speaker recognition. The method, originally suggested by Vapnik, interpreted as a novelty detectors by Tax and Duin. In was used as a classifier. It contains support vectors describing the hypersphere separating the samples. With a minimal radius R, this classifier achieve good performance in finding abnormal samples within the open set test.The third issue is combining support vector machine and Generative Models. Generativ Models such as Hidden Markov Models and Gaussian Mixture Models have been proved to be an efficient way for statistically modeling sequence signals. And the Support Vector Machines seem to be a promising candidate to perform the classification task. To do the combination, probability outputs were extracted from Support Vector Machines.