The effectiveness of ICA-based representation: Application to speech feature extraction for noise robust speaker recognition

In this paper, we present a mathematical derivation demonstrating that feature representation obtained by using the Independent Component Analysis (ICA) is an effective representation for non-Gaussian signals when being both clean and corrupted by Gaussian noise. Our findings are experimentally demonstrated by employing the ICA for speech feature extraction; specifically, the ICA is used to transform the logarithm filter-bank-energies (instead of the DCT which provides MFCC features). The evaluation is presented for a GMM-based speaker identification task on the TIMIT database for clean speech and speech corrupted by white noise. The effectiveness of ICA is analysed individually for signals corresponding to each phoneme. The experimental results show that the ICA-based features can provide significantly better performance than traditional MFCCs and PCA-based features in both clean and noisy speech.