Blind source separation for convolutive mixtures

摘要: Abstract Blind separation of sources is now a well known problem, and various methods have been proposed for instantaneous mixtures, convolutive mixtures narrow-band sources. In this paper, we present adaptive algorithms wide-band signals, modeled by Finite Impulse Response (FIR) filters. The estimate coefficients FIR filters using independence criteria. Under the condition fourth-order white noise, prove in paper that estimation filter may be based on cancellation 4th-order output cross-cumulants. However, simpler only few cross-cumulants or moments, 2nd-order moments same efficacy. Finally, experimental results point out good performance (about -20 dB residual crosstalk) even with nonwhite nonstationary signals like speech signal simulated mixtures. For recorded are not as suggest model realistic enough must improved.