Frequency Domain Approach for Blind Signal Separation of Convolutive Mixed Signals

In this paper a frequency domain approach to blind signal separation (BSS) of convolutive mixed signals is proposed. It is shown that iteratively minimizing the cost function such as the off-diagonal of the cross power spectral (CPS) matrices of the signals at the output of the mixing system is sufficient to identify the separating system at each frequency bin up to a scale and permutation ambiguity. The minimization of the cost function is performed using gradient-based optimization method. Then the inverse of the mixing system is estimated and used to separate the mixed sources. The performance of the proposed algorithm is demonstrated using a mixing of audio signals, BPSK signals, and image signals. The algorithm demonstrates good separation performance and enhanced output audio quality with fast convergence speed relative to the traditional methods.