Blind Source Separation for Convolutive Mixtures Based on Complexity Minimization

ABSTRACT Using algorithmic complexity to perform blind source separation (BSS) was first proposed by Pajunen. This approach presents the advantage of taking the whole signal structure into account to achieve separation, whereas standard ICA-based methods only use either time-correlations or higher order statistics in order to do so. Another advantage of this approach is that no assumptions about the probability distribution of the source signals need to be made. However, although algorithmic complexity based methods have been shown to outperform standard ICA algorithms in the instantaneous BSS case, they haven’t been applied to convolutive BSS to the present date. In this paper, we show that it is also possible to use algorithmic complexity as a separating criterion to perform BSS for convolutive mixtures and suggest a method to do so. Testing the proposed method by computer simulation yielded results which are encouraging in terms of SNR performance.