Novel design of oversampled GDFT filter banks for application to subband based blind source separation

Subband permutation alignment is critical to the successful operation of subband-based blind source separation. The inter-subband correlation information can be exploited to solve the permutation problem. In this paper, a novel design is proposed based on the over-sampled generalized DFT (GDFT) filter banks to further improve performance of the correlation-based method. It employs a model that includes inter-subband correlation as part of the optimization criterion, and relaxes the perfect-reconstruction condition due to an inherent characteristic of the convolutive BSS problem. As shown in simulations, improved performance is achieved in terms of both permutation alignment and overall separation result.