Computer Science Department
School of Computer Science, Carnegie Mellon University


BoB: An Improvisational Music Companion

Belinda Thom

May 2001

Ph.D. Thesis

Keywords: Unsupervised learning, interactive computer music systems, music improvisation, real-time, stochastic processes

In this thesis, I introduce a new melody representation scheme and a machine learning framework that enables customized interaction between a live, improvising musician and the computer. The ultimate intent of these technologies is to provide the infrastructure needed to intimately couple the computer with a musician's all-too-transient improvisations; potential applications range from improvisational exploration to education and musical analysis. I introduce Band-OUT-of-a-Box (BoB) -- a fully realized agent that trades personalized solos with a simulated user in real-time.

Musical improvisation is an ill-defined situation-and user-specific practice whose inherent non-literal basis makes the authoring techniques exploited in other AI/entertainment systems (e.g., interactive characters or stories) less helpful when building an improvisational music companion. A major contribution is BoB's computational melodic improvisation model, which congures itself to its users' unlabeled examples, alleviating some of the burden associated with de ning an appropriate musical aesthetic.

I first describe an abstract perception algorithm that maps short strings of notes onto a mixture model. The components of this model correspond to the various playing modes -- e.g., tonal, intervallic, and directional trends -- that the user employed during various parts of a warmup session. I next describe a crucial technology that closes the perception loop by integrating the learned model's parameters into a stochastic process that, when sampled, can produce sequences that exhibit speci c abstract goals (or playing modes) while seamlessly integrating into the constraints set up by the local environment. These algorithms' musical performances are evaluated by qualitatively exploring their behavior using two different simulations: both the transcriptions of Bebop saxophonist Charlie Parker and jazz violinist Stephane Grappelli. These algorithm's quantitative performance are also assessed using more traditional machine learning techniques.

282 pages

Return to: SCS Technical Report Collection
School of Computer Science homepage

This page maintained by