Generalized Cauchy Distribution (GCD)-based score functions for a fast and flexible subband decomposition ICA

In this paper, we propose a new score function for subband decomposition ICA to improve the algorithm performance. Subband decomposition ICA is known to having better separation performance than frequency-domain ICA. However, it is basically time-domain ICA, with much shorter filters, performed on each subband. In this paper, discussion is focused on an information-maximization based approach which performance depends on the distribution assumption of source signals. We propose the use of a generalized Cauchy distribution as a new distribution assumption to derive a new score function. With the proposed score-function, the performance of subband decomposition ICA algorithm is significantly improved in terms of both SIR and convergence speed.