Doping audio signals for source separation

This work fits in the frames of sparse component analysis (SCA), informed source separation (ISS) and doping watermarking. The SCA relies on a strong hypothesis of sparsity of the sources. In a particular context where the original sources are available (ISS), we make the distributions of the time-frequency coefficients of the sources more sparse, through a doping watermarking that imperceptibly transform the histogram of the coefficients. Using the “sparsified” sources instead of the original ones in a SCA leads to a better estimation of the number of sources and to a more accurate identification of the mixing system.