A frequency-domain normalized multichannel blind deconvolution algorithm for acoustical signals

In this paper, a new frequency-domain normalized MBD algorithm is presented for separating convolutive mixtures of acoustical sources. The proposed algorithm uses unidirectional unmixing filters to avoid backward filtering in gradient terms. The gradient terms are then normalized in the frequency domain. As a result, separation and convergence performances are improved, while whitening effect is relieved greatly. Simulation results with real world recordings demonstrate superior performances of the proposed algorithm.