Dynamics of Beat Perception through a Bayesian Lens

Dynamical systems models of beat perception have successfully reproduced various features of human beat perception, but are tailored to rhythm perception tasks and therefore provide little insight into how beat perception might be conceptually related to other perceptual, cognitive, and motor processes. I argue that rhythm perception can be viewed as dynamic Bayesian inference of a hidden state (rhythm phase, tempo, and meter) based on sensory observations (auditory events) and a model of the probability of an auditory event given the hidden state. When this inference problem is stated formally, solved through approximate Bayesian inference, and simulated, it produces solutions that closely resemble existing dynamical systems models. However, the inference perspective allows us to plug into the unifying theory of predictive processing and the Bayesian brain, and to draw meaningful connections with another dynamic inference process in the brain: tracking the instantaneous state of one's own body and actions. It also provides intelligible cognitive interpretations of abstract model parameters and predicts from first principles how unfulfilled expectations have been shown to warp perceived rhythmic timing.