Convolutive blind source separation based on GDFT filterbanks and pre-determined subband whitening

This paper focuses on the convolutive blind source separation. Confronted by drawbacks of time-domain and frequency-domain approaches, we propose a novel approach for source separation in subband-domain based on the pre-emphasis processing of subband signals. A time-domain algorithm based on the entropy maximization principle, using the natural gradient algorithm for adaptation task, is employed for subband signal separation. Instead of signal whitening based on frame-by-frame linear prediction analysis, we propose a fixed, pre-determined signal whitening scheme in the subbands to improve the separation performance while decreasing artifacts. With less computational complexity and side-effects, the proposed method is experimentally evaluated and shown to be superior to several other subband-based approaches.