Research On Speaker Recognition Technology Based On Support Vector Machine And Multiscale Wavelet Analysis

At present, Speaker Recognition has become a hotspot in authentication identification and artificial intelligence areas. There are important theoretic value and far-reaching practical meaning to resolve the question of speaker recognition in the noising environments. Support Vector Machines is an important theory designed based on the Vapnik-Chervonenkis Dimension in statistical learning theory (SLT) and empirical risk minimization (ERM) principles. As a new machine learning method, SVM shows many good properties over other methods in solving limited samples problems and non-linear high dimension pattern recognition problems, so it is becoming a new study hot area after neural network. At present, SVM is succeeded in the application of some areas, such as model recognition, regression analysis, analysis of time series and cluster analysis, etc. Applying SVM to speaker recognition can resolve some problems of traditional classifier effectively such as over-fitting, bad generalization ability, dimension disaster, etc.A speaker recognition method based on support vector and multi-scale wavelet analysis is proposed and a frame model of it is constructed in this paper. Firstly, Multi-scale wavelet analysis is applied to the process of signal preprocess, based on it, the theory of multi-scale analysis is applied to separate speech and noise, and enhance the speech consequently. Secondly, in the feature extracting phase, MFCC and its difference are derived to be the characteristic parameters and then composed into feature vector sequences based on SVM. Finally, A multi-category SVM algorithm is applied to realize the speaker classification and recognition by making training and testing based on swatches.A real speech library is built in this paper, and a comparative experiment of the recognition performance of SVM is made based on different two speech sets: one of speech de-noised and another without speech de-noising. Experimental results have shown that the recognition rate of speaker recognition system is widely increased by the compositive application of Support Vector Machines and Multi-scale Wavelet Analysis. Furthermore, based on the speech set after wavelet de-noising, some simulated experiments on some questions in the process of speaker recognition are made and the results are given. These simulated experiments mainly focus on the factors of recognition rate, such as amount of characteristic vector, length and dimension of frames, training times, parameters of Kernel function, etc. After that, to increase the recognition rate of speaker recognition system based on the speech set after wavelet de-noising, a speaker recognition method considering gender difference is put forward in allusion to the good classification performance of binary class SVM, and the validity of this method is proved by the simulated experiments. Ultimately, some significant research results are given in this paper, as well as some strong data and positive suggestions for the studies of speaker recognition are offered.