The lack of a formalism for “the problem of music description” results in, among other things: ambiguity in what problem a music description system must address, how it should be evaluated, what criteria define its success, and the paradox that a music description system can reproduce the “ground truth” of a music dataset without attending to the music it contains. To address these issues, we formalize the problem of music description such that all elements of an instance of it are made explicit. This can thus inform the building of a system, and how it should be evaluated in a meaningful way. We provide illustrations of this formalism applied to three examples drawn from the literature.