BLIND SOURCE SEPARATION FOR CONVOLUTIVE MIXTURES USING SPATIALLY RESAMPLED OBSERVATIONS

We propose a new technique for separation of sources from convolutive mixtures based on independent com-ponent analysis (ICA). The method allows coherent processing of all frequencies, in contrast to the tradi-tional treatment of individual frequency bands. The use of an array enables resampling of the signals in such a way that all frequency bands are effectively transformed onto the centre frequency. Subsequent separation is performed “all-bands-in-one”. After res-ampling, a single matrix describes the mixture, allow-ing use of standard ICA algorithms for source separa-tion. The technique is applied to the cocktail-party prob-lem to obtain an initial estimate of the separating para-meters, which may further be processed using cross-talk removal or filtering. Experiments with two sources of speech and a four element microphone array show that the mixing matrix found by ICA is close to the the-oretically predicted, and that 15 dB separation of the sources is achieved. 1