Speaker Identification Based on Multi-reduced SVM

SVM is a novel type of statistical learning methods that has been successfully used in speaker recognition. However, training SVM consumes long computing time and large memory with all training data. This paper proposes a speaker identification method based on multi- reduced support vector machine (MRSVM). MRSVM has two reduction steps. Firstly, speech feature dimensions are reduced by using KL transform, the noise is removed from speech simultaneity. Secondly, speech feature data are selected at boundary of each cluster as SVs by using kernel-based fuzzy clustering technique. Experiment results show that not only the training data, training time and storage can be reduced remarkably, but also the identification accuracy can be improved by the proposed MRSVM compared with other reduced algorithms and the system has better robustness.