In this paper we propose a method for blind signal decomposition that does not require the independence or stationarity of the sources. This method, that we consider a simple instance of non-linear projection pursuit, is based on the possibility of recovering the areas in the time-frequency where the original signals are isolated or almost isolated with the use of suitable quotients of linear combinations of the spectrograms of the mixtures.

We then threshold such quotients according to the value of their imaginary part to prove that the method is theoretically sound under mild assumptions on the mixing matrix and the sources. We study one basic algorithm based on this method. Moreover we propose a practical measure of separation for the sources in a given time frequency representation.

The algorithm has the important feature of estimating the number of sources with two measurements, it then requires n-2 additional measurements to provide a reconstruction of n sources. Experimental results show that the method works even when severalshifted version of the same source are mixed.