ICA-based Rasta-PLP feature for speaker identification

A robust approach that unifies independent component analysis (ICA) feature selection in connection with speaker identification (SI) is proposed. In the feature extraction stage, ICA offers an alternative to discrete cosine transform (DCT), to select relative spectral transform-perceptual linear prediction (RASTA-PLP) feature. ICA 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 speaker identification purpose. The performance of the method is demonstrated with the database prepared in laboratory environment. Experimental results show that the proposed approach is more effective in the corrupted speech case.