STONES: Stochastic Technique for Generating Songs

We propose a novel approach for automatically constructing new songs from a set of given compositions that involves sampling a melody line as well as the corresponding harmonies given by chords. We empirically show that songs generated by our approach are closer to music composed by humans than those of existing methods 1. Composing music is a complex creative process that is not only restricted by harmonic necessities but also by an implicit concept of good music. While music composition algorithms are able to comply with the explicit rules of composing, it is difficult to accord to the implicit concept of a good song. Moreover, this concept may vary between different music styles. We propose to learn the properties of good composed songs from a set of existing ones and use the learned model to construct new songs.