The underdetermined blind audio source separation problem is often addressed in the time-frequency domain by assuming that each time-frequency point is an independently distributed random variable. Other approaches which are not blind assume a more structured model, like the Spectral Gaussian Mixture Models (Spectral-GMMs), thus exploiting statistical diversity of audio sources in the separation process. However, in this last approach, Spectral-GMMs are supposed to be learned from some training signals. In this paper, we propose a new approach for learning Spectral-GMMs of the sources without the need of using training signals. The proposed blind method significantly outperforms state-of-the-art approaches on stereophonic instantaneous music mixtures.