Skip to Main Content
This paper presents a mathematical analysis demonstrating the effectiveness of the signal representation based on independent component analysis (ICA) in the case of non-Gaussian noise corruption. The analysis is based on calculating a mismatch between the distribution of the observed signal represented by a linear model and a reference distribution. The theoretical results lead to a novel ICA-based signal representation technique in which the ICA transformation matrix is estimated based on noise-corrupted signal but not based on clean signal as normal. Our theoretical findings are experimentally demonstrated by employing the proposed feature representation in a GMM-based speaker recognition system. Experimental results show that employment of the proposed ICA-based features can provide significant recognition accuracy improvements over using both the traditional ICA-based features and MFCC features.