Convolutive Blind Source Separation Using Fourier Kalman Filtering

In this paper a frequency domain approach isproposed for convolutive blind source separation (CBSS) ofsignals. The convolutive mixing model when reformulated asa stochastic state-space model and defined in the frequencydomain comes with unknown states and parameters. Thesolution to the problem calls for a dual estimation approachto be applied to recover the original signals. The dualestimation method employed in this paper uses state-spacefrequency domain Kalman filter running a pair of state andparameter filters simultaneously to estimate unknownparameters and states. The performance of the proposedmethod is shown by simulation results and comparisons havebeen made with previous methods.