Study On Speaker Recognition Methods Related To The Text

Speaker Recognition is an important subject of speech processing, It is applied to man-machine interface, ensure public security, military affairs, judicature, and so on. In this thesis, the methods of text dependent speaker recognition are studied, The main works are as follows:(1) According to the discrete time model of speech signal, the feature vector of speaker - LPC cepstrum is extracted, then it is quantized.(2) The statistic learning theory is studied firstly. Based on the theory, support vector machine are studied in details. Most of important problems of support vector machine theory are studied in this thesis. They are the express of machine study problem, empirical risk minimization, the boundary of generalization, structural risk minimization, optimal hyperplane, generalized optimal hyperplane, and optimal hyperplane - support vector machine on high dimension space etc.(3) The method of Hidden Markov Models is studied, including meanings of Hidden Markov Models of speech signal, definition of Markov Chain, category of Hidden Markov Models, how to estimate parameters of Hidden Markov Models and how to realize the algorithm of Hidden Markov Models.(4) Finally, A real speech library is built. The above mentioned methods are simulated and the results of simulation are given.