A Learning-Based Jam Session System that Imitates a Player's Personality Model

This paper describes a jam session system that enables a human player to interplay with virtual players which can imitate the player personality models of various human players. Previous systems have parameters that allow some alteration in the way virtual players react, but these systems cannot imitate human personalities. Our system can obtain three kinds of player personality models from a MIDI recording of a session in which that player participated a reaction model, a phrase model, and a groove model. The reaction model is the characteristic way that a player reacts to other players, and it can be statistically learned from the relationship between the MIDI data of music the player listens to and the MIDI data of music improvised by that player. The phrase model is a set of player's characteristic phrases; it can be acquired through musical segmentation of a MIDI session recording by using Voronoi diagrams on a piano-roll. The groove model is a model that generates onset time deviation; it can be acquired by using a hidden Markov model. Experimental results show that the personality models of any player participating in a guitar trio session can be derived from a MIDI recording of that session.