Complex FastIVA: A Robust Maximum Likelihood Approach of MICA for Convolutive BSS

摘要: We tackle the frequency-domain blind source separation problem in a way to avoid permutation correction. By exploiting facts that frequency components of signal have some dependency and mixing sources is restricted each bin, we apply concept multidimensional independent component analysis propose new algorithm separates groups dependent components. introduce general entropic contrast functions for this corresponding likelihood function with prior models assume circularity complex variables derive fast by applying Newton’s method learning rule. The mixed even very challenging acoustic settings.