Subband-based joint blind source separation for convolutive mixtures employing M-CCA

This paper addresses the source separation problem for convolutive mixtures by employing the joint blind source separation (BSS) technique in the subband domain. The key to the proposed method is to decompose the time-domain mixed signals into subbands to generate the required related multiple data sets for the operation of joint BSS. To reduce the aliasing error after subband decimation, the oversampled generalized DFT filter banks are employed to maintain a sufficient level of correlation between the data sets. A recently proposed correlation optimisation method for the design of filter banks is adopted to enhance the correlation between adjacent subband signals, which leads to further improved separation results in terms of both signal to interference ratio and subband permutation alignment.