Optimized dyadic sorting for solving the permutation ambiguity in acoustic blind source separation

In this paper, we propose a modification to dyadic sorting scheme used for the permutation problemin convolutive blind source separation. In the frequency domain, the problem of separation of sources can be reduced to multiple instantaneous problems, which are easily solvable using independent component analysis. However, this simplified method leads to the problem of correctly aligning and scaling of the single frequency bins. These ambiguities need to be solved before the transformation to the time domain, as otherwise the separation process will fail. In this paper we combine dyadic sorting with an optimized way of calculation of correlation coefficients by using spectral summation. The improved performance will we shown on real world examples.