Blind Separation of Convolutive Mixtures using Nonstationarity and Fractional Lower Order Statistics (FLOS): Application to Audio Signals

In this paper, we introduce new time-varying fractional spectral matrices to exploit both the nonstationarity and heavy-tailed sources properties for blind separation of convolutive audio mixtures. We define these spectrum matrices, that are different for various delays, using fractional lower order statistics (FLOS) of data. Similar to the second order statistics (SOS) based approaches, we maximize the sources independence by jointly diagonalizing these fractional matrices spectrum of the reconstructed signals using a mutual information criterion. A set of experiments using audio signals and real impulse response of acoustic room are designed to verify the usefulness of the proposed method