The Efficiency of ICA-based Representation Analysis: Application to Speech Feature Extraction

This paper is focused on the theoretical analysis of the efficiency of Independent component analysis (ICA) based non-Gaussian signal representations corrupted by different types of noises. We first present a mathematical derivation demonstrating that ICA-based signal representations, in which the ICA transformation matrix was derived on clean non-Gaussian signals (denoted by ICA_c) are efficient representations than DCT for signals when being both clean and corrupted by Gaussian noise, while it may not have better performance for the non-Gaussian noises corrupted signals. The analysis also demonstrates that to obtain efficient representations of non-Gaussian corrupted signals, the ICA transformation matrix should be obtained from noise corrupted signal (denoted by ICA_n). The ICA_n-based features can provide significant recognition accuracy improvements to non- Gaussian corrupted signals over both the ICA_c-based features and MFCC features. 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 Gaussian noises and non- Gaussian noises.