SVM is a novel type of statistical learning method that has been successfully used in speaker recognition. However, training SVM consumes long computing time and large storage space with all training examples. This paper proposes an improved sparse least-squares support vector machine (LS-SVM) for speaker identification. Firstly KPCA is exploited to reduce the dimension of input vectors and to denoise speech signal by extracting the nonlinear principal components of feature vectors. Since LS-SVM simplifies the computation by solving a set of linear equations instead of the quadratic programming problems involved by the standard SVM, LS-SVM classification algorithm has been run in our identification system. However before training samples, we have used pruning method to reduce the number of training samples which have been preprocessed by KPCA without discounting the generalization performance. A number of experimental results illustrate that the proposed method shows faster speed and greater accuracy with less storage than other models.
展开▼
