Single autoterms selection for blind source separation in time-frequency plane

We address in this paper a method to perform blind separation of over-determined instantaneous linear mixtures of non-stationary sources. Our contribution is focused on the identification of single source autoterms which correspond to diagonal Spatial Time-Frequency Distribution of the sources vector with only one non-zero diagonal entry. These latter matrices are processed into an iterative joint-diagonalization scheme which provides an estimation of the mixing matrix.