The Study Of Text-independent Speaker Recognition Alogrithm Based On SVM

Due to its special merits of flexibility, economy and accuracy, speaker recognition technology has a broad application prospect in biometrics security field, especially in E-commerce, Criminal identification, Information Security ect . Speaker recognition techniques have performed well under ideal conditions; however, there are still many problems when we want to apply speaker recognition into practical applications. One important reason is the long computational time when training a speaker model or testing an utterance. This makes real-time implementation very hard and expensive. In order to improve train and recognition speed without deteriorating recognition performance. The support vector machine was used for recognition and two speaker recognition methods were proposed in this paper:1) Speaker identification based on multi-reduced SVMThrough the study of reduced support vector machine, we find that in RSVM the support vectors are selected randomly. We also know that in speaker recognition, the training samples are very lager. In order to overcome these shortcomings, a speaker identification approach based on the multi-reduced SVM (MRSVM) was proposed in this paper. First the entropy-based feature selection approach was exploited to reduce the dimension of the input vectors and improve the performance of the clustering .And then the kernel-based possibilistic C-means(KPCM) clustering algorithm had been run on the selected samples to give out a series of representative vectors which were applied to train RSVM as support vectors in high space. By doing so, the storage and the amounts of training data were reduced respectively. The experimental results show that the identification speed is improved further more compared with the standard SVM.2) Speaker recognition based on improved Particle Swarm OptimizationAiming at the shortcoming of Particle Swam Optimization (PSO) which was easilyrelapsed into local extremum, an improved PSO was proposed in this paper. In this approach, we applied the evolution speed factor as the Trigger conditions to stochastically disturb the local optimal solution. The improved PSO algorithm could not only improve extraordinarily the convergence velocity in the evolutionary optimization, but also can adjust the balance between global and local exploration suitably. Then a speaker identification approach using this improved algorithm to train SVM was presented. The SVM can receive the optimal hyperplane with less support vectors by the improved PSO, and then the training samples are reduced and the recognition speed is improved.