Dimension reduction of feature vectors using WPCA for robust speaker identification system

Abstract

Speaker identification based on speech signal has been receiving enhanced attention from the research community. In this context the effect of dimension reduction of feature vectors using Principal Component Analysis (PCA) and Weighted Principal Component Analysis (WPCA) are compared for speaker identification in a noisy environment. MFCC feature vectors are used as original features and their dimension is reduced by PCA and WPCA techniques and then evaluated by GMM classifier. Speaker identification rate is calculated under different SNR to test the robustness of the speaker identification system. In low SNR, the speaker identification rate becomes double after reducing the dimension of feature vectors by 50% as compared to original one. The performance of WPCA is 10% to 20% better than PCA under different SNR.