Feature Selection by Independent Component Analysis for Robust Speaker Verification

A robust approach that unifies independent component analysis (ICA) subspace feature selection in connection with the speaker verification (SV) is proposed. ICA subspace provides statistically independent basis that spans the input space of corrupted speech, then the selected independent components are applied to a vector quantizer (VQ) for SV purpose. The Euclidean distance in the feature space is kept invariant by using ICA and is also used in the VQ based SV system as a matching choice. In the feature selection stage, a batch-mode FastICA algorithm and two adaptive algorithms EGLD-ICA and Pearson-ICA are employed for two-microphone case. As a result, the selected features provide a lower classification error and a better generalization in real environments. The performance of the approach is demonstrated with YOHO database [8] in cocktail party effect and ambient noise cases.