This paper addresses the link between the size of the dictionary in overcomplete decompositions of signals and the rate-distortion properties when such decompositions are used for audio coding. We have performed several experiments with sets of nested dictionaries showing that very redundant shift-invariant and multi-scale dictionaries have a clear benefit at low bit-rates ; however for very low distortion a lot of atoms have to be encoded, in these cases orthogonal transforms such as the MDCT give better results.