Blind Source Separation in nonminimum-phase systems based on filter decomposition

This paper focuses on the causality problem in the task of Blind Source Separation (BSS) of speech signals in nonminimum-phase mixing channels. We propose a new algorithm for solving this problem using filter decomposition approach. Our proposed algorithm uses an integrated cost function in which independence criterion is defined in frequency-domain. The parameters of demixing system are derived in time-domain, so the algorithm has the benefits of both time and frequency-domain approaches. Compared to the previous work in this framework, our proposed algorithm is the extension of filter decomposition idea in multi-channel blind deconvolution to the problem of blind source separation of speech signals. The proposed method is capable of dealing with both minimum-phase and nonminimum-phase mixing situations. Simulation results show considerable improvement in separating speech signals specially when the mixing system is nonminimum-phase.