An integrated method for blind separation and dereverberation of convolutive audio mixtures

This paper proposes an integrated method for the blind separation and dereverberation of convolutive audio mixtures. The proposed method is based on multichannel blind deconvolution in the frequency domain. Significant points to be emphasized are as follows: (1) The objective function for optimizing the deconvolution system is derived based on a time-varying all-pole model of source signals, which was proven to be effective for single source dereverberation. This provides the proposed method with the capacity for both separation and dereverberation. (2) An efficient optimization algorithm is developed. This algorithm is realized by decomposing the deconvolution system into an instantaneous separation part and a multichannel auto-regressive part. Illustrative experimental results with an RT₆₀ of 0.6 seconds are reported, where the proposed method showed superiority over a conventional frequency-domain blind separation method.