A Novel Multi-reduced SVM Approach for Speaker Recognition

To overcome the vast computation of standard SVM, a novel multi-reduced SVM method for speaker recognition is proposed in this paper. The proposed method consists of three parts. Firstly the entropy-based feature selection approach is exploited to reduce the dimension of the input vectors by extracting the important feature attributes, in which the performance of the clustering is improved. Secondly the kernel-based possibilistic C-means (KPCM) clustering algorithm has been run on the selected samples to give out a series of clustering centers, which can represent better the clusters they belong to in high space. Finally, these clustering centers are applied to train RSVM as support vectors. By doing so, we can ensure that the loss of information is minimum. The experimental results show that the training time and storage can be reduced remarkably without deteriorating the recognition performance by the proposed method compared with other reduced algorithms.