Underdetermined source separation is often carried out by modeling time-frequency source coefficients via a fixed sparse prior. This approach fails when the number of active sources in one time-frequency bin is larger than the number of channels or when active sources lie on both sides of an inactive source. In this article, we partially address these issues by modeling time-frequency source coefficients via Gaussian priors with free variances. We study the resulting maximum likelihood criterion and derive a fast non-iterative optimization algorithm that finds the global minimum. We show that this algorithm outperforms state-of-the- art approaches over stereo instantaneous speech mixtures.