Frequency-Domain Normalized Multichannel Blind Deconvolution for Convolutive Speech Mixtures: Modifications and Properties

The normalized multichannel blind deconvolution algorithm and its modifications are described in detail. The proposed algorithm is shown to have fast uniform convergence and improved separation performances for nonstationary sources. In addition, the local stability problem, which limits the separation performance of the algorithm, is investigated in conjunction with source-microphone geometry. Then local stability is treated by decomposing the mixed signals into into direct and reverberant parts. Cascading a beamformer and blind source separation with causal filters is proposed as a practical method to improve local stability. Simulation results using real-world recordings confirm the theoretical expectations.