Discriminative feature extraction for language identification

In this paper we propose a discriminative feature extraction method, DFE, to address the increasing number of features in language identification (LID) tasks. Similar to linear discriminant analysis (LDA), it extracts the most discriminative features through the maximization of an “approximated” mutual information I(C; Y ) between the class labels C and the projected data Y. Compared with other feature extraction methods, experiments done on the CallFriend corpus shows DFE could handle high-dimensional dataset with ease. Furthermore, this feature extraction shows improvements on the LID task over standard feature extraction methods (LDA and principal components analysis).