A Clustering Algorithm for Recombinant Jazz Improvisations

Music, one of the most structurally analyzed forms of human creativity, provides an opportune platform for computer simulation of human artistic choice. This thesis addresses the question of how well a computer model can capture and imitate the improvisational style of a jazz soloist. How closely can improvisational style be approximated by a set of rules? Can a computer program write music that, even to the trained ear, is indistinguishable from a piece improvised by a well-known player? We discuss computer models for jazz improvisation and introduce a new system, Recombinant Improvisations from Jazz Riffs (Riff Jr.), based on Hidden Markov Models, the global structure of jazz solos, and a clustering algorithm. Our method represents improvements largely because of attention paid to the full structure of improvisations. To verify the effectiveness of our program, we tested whether listeners could tell the difference between human solos and computer improvisations. In a survey asking subjects to identify which of four solos were by Charlie Parker and which were by Riff Jr., only 45 percent of answers among 120 people were correct, and less than 5 percent identified all four correctly.