A multi-objective optimization model for music styles

Style recognition is one of the problems mostly faced by Computational Intelligence techniques. Most of them were defined ad-hoc for a specific music genre and so not generalizable and applicable to any style. A music style, both of soloists performer and of musical collectives, is the result of aesthetic goals, i.e., experience and preferences, functional rules, i.e., rules used to produce music, and external influence, i.e., the choices depending by the simultaneous presence of other musicians. We propose a new model of style, defined in terms of a multi-objective problem, where the objective is to minimize the distance between the style of each musician and the stylistic features derived by other musicians. Such a model is general since it is applicable to any type of style. We also propose a new approach for both recognition and automatic composition of styles based on such a model, which exploits a machine learning recognizer and a splicing composer. To assess the effectiveness and the generalization capability of our system we performed several tests using a large set of Jazz transcriptions and a corpus of 4-voice music by J. S. Bach. We show that our classifier is able to achieve a recognition accuracy of 97.1%. With regard to the composition process, we measured the capability of our system to capture both aesthetic goals by collecting subjective perceptions from domain experts, and functional rules by computing the average percentage of (1) typical harmonic progressions in the Jazz music produced and (2) forbidden exceptions, which occur in the 4-voice music, produced.