CONSTRAINED MUSIC GENERATION USING MODEL-CHECKING
Résumé
Numerous works have tackled the problem of style mod- eling: learn some implicit notion of style from an ensemble of musical sequences and generate new content resembling this set of examples. This paper deals with the additional problem of introducing high-level constraints into such applications, that is, structured generation: generate new music in the style of existing musical data, satisfying a given temporal structure.
We propose an original approach to solve this issue through the use of model-checking. It is applied here to the Factor Oracle, an automaton that has already been used by composers for music generation, but currently lacks high-level control capacities and therefore structure. FO captures the sequential structure of symbolic or signal-level musical inputs. Following an abstraction of the automa- ton structure, the specifications are converted to temporal- logic formulae and solutions are efficiently searched for by means of external model-checking tools. This search is done in a backward manner, allowing for the retrieval of optimal solutions.
A Python implementation of the system and musical examples, e.g. for the extraction of chord sequences from a piece, are provided.
Origine : Fichiers éditeurs autorisés sur une archive ouverte