Separation of convolutive mixtures in the frequency-domain using only two frequency bins

In this paper we show that a convolutive mixture of statistically independent signals (sources) can be separated in the frequency-domain using only two frequency bins. The idea deals with finding a 2 × 2 invertible matrix which relates the frequency-domain sources and the temporal sources. This matrix exists when the Fourier transform of the observations is computed using zero-padding and sliding windows with delay of one point. We also propose a blind approach where a well-known algorithm is used and the amplitude/permutation indeterminacy is removed using a strategy based on the statistically independence of the temporal sources.