Particle Filters for Efficient Meter Tracking with Dynamic Bayesian Networks

Recent approaches in meter tracking have successfully applied Bayesian models. While the proposed models can be adapted to different musical styles, the applicability of these flexible methods so far is limited because the application of exact inference is computationally demanding. More efficient approximate inference algorithms using particle filters (PF) can be developed to overcome this limitation. In this paper, we assume that the type of meter of a piece is known, and use this knowledge to simplify an existing Bayesian model with the goal of incorporating a more diverse observation model. We then propose Particle Filter based inference schemes for both the original model and the simplification. We compare the results obtained from exact and approximate inference in terms of meter tracking accuracy as well as in terms of computational demands. Evaluations are performed using corpora of Carnatic music from India and a collection of Ballroom dances. We document that the approximatemethods perform similar to exact inference, at a lower computational cost. Furthermore, we show that the inference schemes remain accurate for long and full length recordings in Carnatic music.