Blind separation of wide-band sources in the frequency domain
Valérie Capdevielle, Christine Servière, Jean‐Louis Lacoume · 2002
Conventional antenna array processing techniques are based on the use of second order statistics but rest on restrictive assumptions. Thus, when a priori information about the propagation or the geometry of the array are hardly available, the model can be generalized to a blind sources separation model. It supposes the statistical independence of the sources and their non-Gaussianity. We focus on the generalization of the sources separation problem to convolutive mixtures of wide-band sources in the frequency domain. As the N-point discrete Fourier transform of signals generally tends to be Gaussian when N tends to infinity according to the central limit theorem, we replace the independence criterion of the sources with the independence of a specific MA filtering of their DFT. The result is then an instantaneous complex mixture in time-domain and the new sources are proved to be non-Gaussian.