In this contribution, we relate the question of discernability of sound structures to the properties of the underlying analysis tools. In particular, we argue, that classical tools that are mainly used in sound processing and lead to features as prominent as the MFCC may be replaced by more accurate methods that are based on rather recent mathematical signal processing tools. In particular, we focus on adaptive representations that lend themselves to efficient computation and, on the other hand, on sparsity-promoting methods which are able to adapt to the structures present in a particular signal class.