This master thesis is on the use of semi-supervised NMF for audio sources separation, and in particular speech separation. The main contribution of this work is a speech separation method based on an adapta- tion to an unknown speaker of a prior training performed with a known speaker. First, an overview of the speech separation by semi-supervised NMF methods - and their current limits in the case of an unknown speaker - is presented, as well as a state-of-the-art of improvements proposed until now. After discussing those different approaches, a new adaptation method to an unknown speaker of the training performed with a known speaker is presented, as well as several constraints aimed at improving the separation quality. Finally, the proposed adaptation model and the different constraints are evaluated and compared to the results obtained without speaker adaptation.