We propose in this paper a unique method to separate sources that may have different statistical properties, in the case of FIR convolutive mixtures. No constraint is necessary on the source statistics (i.i.d variables, Gaussian sources or temporally correlated sources..), nor on the number of each type of sources. On the contrary of previous works, no assumption of overdeter-mined mixtures is used. It relies on joint block-diagonalization of correlation matrices of some appropriate variables called differential complex signals, which are introduced in the paper.
展开▼
