We consider the source extraction problem for stereo instantaneous musical mixtures with more than two sources. We prove that usual separation methods based only on spatial diversity have performance limitations when the sources overlap in the time-frequency plane. We propose a new separation scheme combining spatial diversity and structured source priors. We present possible priors based on nonlinear Independent Subspace Analysis (ISA) and Hidden Markov Models (HMM), whose parameters are learnt on solo musical excerpts. We show with an example that they actually improve the separation performance.