Sparsity has been shown to be very useful in blind source separation. However, in most cases the sources of interest are not sparse in their current domain and are traditionally sparsified using a predefined transform or a learned dictionary. In this paper, we address the case where the underlying sparse domains of the sources are not available and propose a solution via fusing the dictionary learning into the source separation. In the proposed method, a local dictionary is learned for each source along with separation and denoising of the sources. This iterative procedure adapts the dictionaries to the corresponding sources which consequently improves the quality of source separation. The results of our experiments are promising and confirm the strength of the proposed approach.