Abstract : Magnetoencephalography (MEG) and electroen-cephalography (EEG) source localization is a challenging ill-posed problem. To identify an appropriate solution out of an infinite set of possible candidates, the problem requires setting certain constraints depending on the assumptions or a priori knowledge about the source distribution. Different constraints have been proposed so far, including those that impose sparsity on the source reconstruction in both standard and time-frequency domains. Source localization in the time-frequency domain has already been investigated using Gabor dictionary in both a convex (TF-MxNE) and non-convex way (Iterative Reweighted TF-MxNE). The iterative reweighted (ir)TF-MxNE solver has been shown to outperform TF-MxNE in both source recovery and amplitude bias. However, the choice of an optimal dictionary remains unsolved. Due to a mixture of signals, i.e. short transient signals (right after the stimulus onset) and slower brain waves, the choice of a single dictionary explaining simultaneously both signals types in a sparse way is difficult. In this work, we introduce a method to improve the source estimation relying on a multi-scale dictionary, i.e. multiple dictionaries with different scales concatenated to fit short transients and slow waves at the same time. We compare our results with irTF-MxNE on realistic simulation, then we use somatosensory data to demonstrate the benefits of the approach in terms of reduced leakage (time courses mixture), temporal smoothness and detection of both signals types.