A robust method to count, locate and separate audio sources in a multichannel underdetermined mixture

We propose a method to count and estimate the mixing directions and the sources in an underdetermined multichannel mixture. Like DUET-type methods, the approach is based on the hypothesis that the sources have time-frequency representations with limited overlap. However, instead of assuming essentially disjoint representations, we only assume that, in the neighbourhood of some time-frequency points, only one source contributes to the mixture: such time-frequency points can provide robust local estimates of the corresponding source direction. At the core of our contribution is a local confidence measure --inspired by the work of Deville on TIFROM-- which detect the time-frequency regions where such a robust information is available. A clustering algorithm called DEMIX is proposed to merge the information from all time-frequency regions according to their confidence level. Two variants are proposed to treat instantaneous and anechoic mixtures. In the latter case, to overcome the intrinsic ambiguities of phase unwrapping as met with DUET, we propose a technique similar to GCC-PHAT to estimate time-delay parameters from phase differences between time-frequency representations of different channels. The resulting method is shown to be robust in conditions where all DUET-like comparable methods fail: a) when time-delays largely exceed one sample; b) when the source directions are very close. As an example, experiments show that, in more than 65% of the tested stereophonic mixtures of six speech sources, DEMIX-Anechoic correctly estimates the number of sources and outperforms DUET in the accuracy, providing a distance error 10 times lower.