Second order impropriety based complex-valued algorithm for frequency-domain blind separation of convolutive speech mixtures

The performance of the complex-valued blind source separation (BSS) is studied in the frequency domain approach to separate convolutive speech mixtures. In this context, the strong uncorrelating transform (SUT) and complex maximization of non-Gaussianity (CMN) do not produce satisfactory separation results since their assumptions about the independence among the frequency-domain complex-valued sources and the different diagonal elements of the pseudo-covariance of those sources are not met at each frequency bin. The proposed strong second order statistics (SSOS) algorithm exploits the second order impropriety of the frequency-domain complex-valued sources with the assumption that the complex-valued sources are improper and uncorrelated, and can well separate the mixtures at about 50% of frequency bins, outperforming SUT and CMN. Thus, it is promising to recover the time-domain speech sources by combing SSOS and the following indeterminacy correction in the frequency domain approach to separate convolutive speech mixtures.