SUBBAND-BASEDBLINDSIGNALPROCESSINGFOR SOURCE SEPARATIONINCONVOLUTIVEMIXTURESOF SPEECH

This paper describes ahighly practical blind signal separation (BSS) scheme operating onsubband domain data toblindly segregate convolutive mixtures ofspeech. Theproposed method relies onspatiotemporal separation carried outinthetimedomain byusing amultichannel blind deconvolution (MBD)algorithm that enforces separation byentropy maximization through thepopular natural gradient algorithm (NGA). Numerical experiments withbinaural impulse responses affirm thevalidity andillustrate thepractical appeal ofthe presented technique evenfordifficult speech separation setups. Index Terms-Subband filtering, blind source separation, multichannel blind deconvolution, convolutive speech mixtures.