A multiresolution approach to blind separation of speech signals in a reverberant environment

The performance of existing blind speech separation methods is limited in a realistic reverberant environment, where a need for long un-mixing filters is imperative. We first show how these methods suffer while trying to balance the competing objectives of frequency-domain permutation alignment and spectral resolution. We then propose a multistage multiresolution algorithm, which aligns the un-mixing filter permutations over the whole frequency band without sacrificing spectral resolution. We perform experiments in both real and simulated reverberant environments, and obtain improved separation results that are comparable to the ideal benchmark obtained by aligning the permutations using prior knowledge of the mixing filters.