AVATAR Machine Learning Improvising Software

Screen Shot 2020-12-13 at 4.02.19 AMAt IUPUI’s Tavel Lab, Jason Palamara and I have been collaborating on the production of self-driving musical software that listens to a vibraphone player and improvises along side. This performance shows a performance that was featured at the Summer Intensive in Contemporary Performance Practice (SICPP), July 31, 2020. The Avatar program is a machine-learning-enabled “choice engine” which provides a dynamically sensitive duet while listening to live vibraphone performances. The initial version is geared for use with a vibraphone, with additional instruments soon to follow. Using this system, the musician performs improvisations on the vibraphone while the software listens, closely following the vibraphone performance. The package employs a Markov-chain model culled from Scott Deal’s improvisations. This mindfile database allows the software to generate novel content based on Scott Deal’s style. While the Markov transition database provides note-to-note transitions, the AvatarPlayer makes use of this data in several ways. Throughout a performance, the AvatarPlayer cycles through five playback behaviors (favor repetition, favor novelty, favor four notes, favor chords, and favor phrases), all of which make use of the database differently.
Music by Scott Deal, Software Design by Jason Palamara