Fast convolutive blind speech separation via subband adaptation

In this paper, we consider the problem of blind source separation (BSS) applied to speech signals. Due to reverberation, BSS in the time domain is usually expensive in terms of computations. We propose in this paper a subband BSS system based on the use of adaptive feedback de-mixing networks in an oversampled uniform DFT filter bank structure. We show that the computational cost can be significantly decreased if BSS is carried out in subbands due to the possibility of reducing the sampling rate. Experiments with real speech signals, conducted with two-input two-output BSS systems using oversampled 32-subband and fullband adaptation, indicate that separation quality and distortion are similar for both systems. However, the proposed subband system is more than 10 times computationally faster than the fullband one.