Description-Based Design of Melodies

We introduce a novel paradigm for creating musical objects, based on the combination of a machine-learning and a combinatorial algorithm. In this scheme users associate freely subjective descriptions (tags) to musical objects. A machine-learning component continuously learns a mapping between these tags and a set of technical features. Thanks to a combinatorial generator, the user can then reuse these tags to modify other objects, in an incremental manner. In this scheme, to the traditional construction or programming task is substituted a description task integrated with an interactive naming game. We describe the paradigm and its application to the construction of simple monophonic melodies. We show that the approach allows users to create “interesting” melodies of various types without requiring any form of explicit programming. This approach to music composition lessens the need for technical skills from the composer, while exploiting fully his capacity to express and manipulate consistent subjective judgments.