Recordings of audio often show undesirable alterations, mostly the presence of noise or the corruption of short parts. Clipping, or saturation, is one of such alterations. Several techniques have been developed in order to attempt the reversal of this corruption, achieving good but perfectible results. One of these techniques, developed in the METISS project-team, involves the use of sparse representations, a popular model in signal processing. The principle of sparse representations is to describe a high-dimensional data vector as a linear combination of a few prototype vectors, called atoms, selected from a large corpus called the dictionary. Building upon this technique, the aim of this internship is to define if and how can machine learning be applied on the dictionary in order to further enhance the results: what to learn on, with what learning algorithm, and with what kind of signals does it work?