SVM-UBM Based Automatic Language Identification Using a Vowel-guided Segmentation

As powerful theoretical and computational tools, support vector machines (SVMs) have been widely used in pattern classification of many areas. A key issue of applying SVMs to language identification of speech signals is to find a SVM kernel that compares a sequence of feature vectors with others efficiently. In this paper, we introduce a sequence kernel used in language identification, and develop a Gaussian Mixture Model to do the sequence mapping task, which maps a variable length sequence of vectors to a fixed dimensional space. Experiment results demonstrate that the new system not only yields performance superior to those of a GMM classifier but also outperforms the system using Generalized Linear Discriminant Sequence (GLDS) Kernel.