Learning Models for Interactive Melodic Improvisation

Belinda Thom School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213, U.S.A. Abstract. This research addresses the problem of the computer interacting with a live, improvising musician in the jazz/blues setting. We introduce BoB, a model of improvisation that enables the computer to trade solos with a musician in an adaptive, user-speci c manner. We develop unsupervised learning methods for autonomously customizing the model via improvised examples and demonstrate the powerful musical abstractions that emerge when applied to Charlie Parker's Mohawk improvisations. Our key technical contribution is the development of an architecture that naturally enables unsupervised learned knowledge, perception and generation to be tightly coupled.