We present a new domain for unsupervised learning: automatically customizing the computer to a specific melodic performer by merely listening to them improvise. We also describe BoB, a system that trades customized real-time solos with a specific musician. We develop a probabilistic mixture model, derived from the multinomial distribution, for the clustering and generation of variable sample-sized his-tograms. With this model, bars of a solo are clustered via the pitch-classes contained therein, adding a new dimension to the problem; the need to learn from sparse histograms. With synthetic data, we quantify the feasibility of handling this issue, and qualitatively demonstrate that our approach discovers powerful musical abstractions when trained on saxaphonist Charlie Parker.
展开▼
