Drum'n'Bayes: On-Line Variational Inference for Beat Tracking and Rhythm Recognition
Charles Fox, Iead Rezek and Stephen J. Roberts
In: roceedings of the the International Computer Music Conference(2007).
It is useful for music perception and automated accompaniment systems to perceive a music stream as a series
of bars containing beats. We present a proof-of-concept
implementation of a Variational Bayesian (VB) system
for simultaneous beat tracking and rhythm pattern recognition in the domain of semi-improvised music. This is
music which consists mostly of known bar-long rhythm
patterns in an improvised order, and with occasional unknown patterns. We assume that a lower-level component is available to detect and classify onsets. The system
uses Bayesian network fragments representing individual
bars and infers beat positions within them. Model posteriors provide principled model competition, and the system may be seen providing a Bayesian rationale for agent-based and blackboard systems. The psychological notion
of priming is used to instantiate new candidate models.