A Novel Reduction Method for Text-Independent Speaker Identification

SVM is a novel statistical learning method that has been successfully applied in speaker recognition. However, Extractive feature vectors from the speech are overlapped and noisy is included in the original data space, these problems can lead to experience difficulties, training complication during training SVM, and the result will be reduced during the recognition phase. In this paper, a novel method is proposed to reduce the noise and input vectors of the SVM. Firstly data dimensions are reduced and noise is removed by using PCA transform, secondly feature data are selected at boundary of each cluster as SVs by using Kernel-based fuzzy clustering technique. The training data, time and storage can be reduced remarkably compared with traditional SVM; the speaker identification system based on our proposed reduced support vector machine (RSVM) has better robustness compared with other reduced algorithms.