Belinda Thom, Carnegie Mellon University
We present a new and exciting domain for unsupervised learning: automatically customizing the computer to a specific melodic performer by listening to them improvise. We also describe our system BoB, which performs this task in the context of real-time solo trading. We develop a probabilistic mixture model of variable-sized multinomials and a procedure that uses this model to learn how to perceive/generate variable-sized histograms. This model is used to cluster bars of improvisation based on the nominal pitch-classes played therein, which adds a new dimension to the problem: the need to learn from sparse data. With simulated data, we show that useful results can be learned with sparse histograms. In BoB, we show that this approach enables powerful musical abstractions to emerge on multiple levels for bebop saxaphonist Charlie Parker.