By François Pachet.
Computer Music Journal, 33(4), pp. 56–68, Winter 2009.
Read the paper (PDF)
Publisher record
About the paper
This paper introduces a way for musicians to steer melodies using words they choose themselves. A user labels example melodies with descriptions such as “jumpy” or “tonal.” The system learns how that user applies each word, then uses it to guide changes to a selected melody.
Abstract summary
For each descriptive tag, a classifier learns from symbolic features of the labeled melodies. Starting with an existing melody, the system searches for nearby variations that score higher or lower for the selected tag. Users can label the new results to refine what their words mean to the system. The paper demonstrates progressive changes in tonality, length, and Brownian melodic motion.
Related work
The approach grows out of the work on reflexive interaction described on the Continuator project page. See Flow Machines for later research on helping musicians steer composition.