Synthesizing musical accompaniments with Bayesian belief networks

This paper discusses recent work in creating a computer program that plays the role of a sensitive musical accompanist. An accompanist must synthesize a number of different sources of information including a real-time analysis of the soloist's acoustic signal, an understanding of the timing relationships represented in the musical score, the interpretation of the soloist learned through rehearsals, and musical constraints on the way in which the accompaniment can be played. A probabilistic framework is presented in which all of these knowledge sources can be represented and learned from actual data. This model then forms the basis of an approach to musical accompaniment using the machinery of Bayesian belief networks. A demonstration is provided from J.S. Bach's Cantata 12.