Feature concatenation for speaker identification

The use of feature vectors obtained by concatenation of different features for text independent speaker identification from clean and telephone speech is studied. The composite feature vectors are examined with GMM and VQ models used to classify speakers. Linear discriminant analysis (LDA), a statistical tool designed to select a reduced set of features for best classification, is applied to enhance performance. The use of LDA for reducing the size of composite feature vector was found satisfactory for clean speech but not for telephone speech. On the other hand, using LDA in the not conventional manner — as a nonsingular transformation (i.e. without size reduction) — improved the performance of composite features in both clean and the telephone speaker identification experiments.