A new permutation control method for frequency domain BSS

This paper introduces a new frequency domain blind source separation algorithm: Inter-frequency Correlation with Microphone Diversity (ICMD). Here, we consider using different sets of microphones where in each set the number of microphones and sources are equal. In the frequency domain, cascaded ICA initialization (CII) is used, where the separation matrix of one bin is used to initialize the ICA iterations of the next. CII greatly reduces the number of permutation changes in successive bins. However, for a given microphone set, it is not uncommon that ICA will fail to separate some bins, thus defeating CII. This problem is addressed as follows. 1) In addition to CII the inter-frequency correlation matrix of the separated signals is used to align permutations in successive frequency bins. 2) The condition number of this matrix is monitored to determine if separation has failed for the current bin and microphone set. 3) If so, an alternate set with better separation is selected and again, inter-frequency correlation is used to align the permutations of the new set of microphones with the old. Results show a marked improvement in separation when there are three or more sources.