Jazz piano trio synthesizing system based on HMM and DNN

In this paper, we discuss a computational model of an automatic jazz session, which is a statistically trainable. Moreover, we describe a jazz piano trio synthesizing system that was developed to validate our model. Most previous mathematical models of jazz session systems require heuristic rules and human labeling of training data to estimate the musical intention of human players in order to generate accompaniment performances. In contrast, our goal is to statistically learn the relationship between a piano, a bass, and a drum player from performance MIDI data as well as information contained in lead sheets, for instance tonic key and chord progression. Our system can generate the performance data of bass and drums from only piano MIDI input, by learning the interrelationship of their performances and time series characteristics of the three involved instruments. The experimental results show that the proposed system can learn the relationship between the instruments and generate jazz piano trio MIDI output from only piano input.