Enhancements of SVM Speaker Recognition by Dynamic Time Wrapping

Abstract

This paper proposes an approach using dynamic time wrapping (DTW) to improve the classification performance of the SVM separation hyperplane. The presented method by incorporating the distance information derived from DTW template matching calculations into SVM separation hyperplane training will be able to effectively control the balance between mar-gin maximization and the amount of misclassifications, and therefore the recognition accuracy of the SVM classifier on speaker recognition will further be increased. Experimental results demonstrated the effectiveness and efficiency of the developed approach.