An adaptive short-time frequency domain algorithm for blind separation of nonstationary convolved mixtures

We present a frequency domain algorithm derived for windowed-adaptive blind separation of convolved sources. Signal separation (filtering) is performed in short-time-windowed-frequency domain in terms of a finite filter length L obtaining faster convergence and better performance compared with the strictly time domain algorithms. In order to avoid the whitening effect the recurrent neural network, similar to the one proposed by Back and Tsoi (1994), is employed. A statistical independence test is done in time domain in order to determine the relative time-varying effect and solve the permutation indeterminacy problem. Corrections of the learning rules are introduced, which show to improve separation performance significantly. Additionally, the results developed by Amari et al. (2000) for the instantaneous mixtures are applied making learning equations computationally more efficient. To resolve the permutation problems the neural network outputs algorithm developed by Markowitz and Szu (1999) is applied.