| dc.description.abstract | Rhythms are central in human speech and music, with the tradition of using computational models to understand the unique human talent for flexibly processing and synchronising movement with auditory rhythms dating back to the early history of cognitive science. Previous approaches to developing computational cognitive models of rhythm processing are either based on discrete (symbolic) processing and representations, simulating long-term learning of rhythms within a musical culture; or continuous (sub-symbolic) processing, simulating short-term dynamics of synchronisation to musical rhythms. The disconnect between these existing models limits our psychological understanding of rhythm processing, particularly at the intersection of two core cognitive processes: enculturation and entrainment. Motivated by the predictive processing framework, the present research characterises rhythmic expectation and synchronisation through a series of formal inference problems, and incrementally defines a hierarchical probabilistic generative model capable of performing continuous Bayesian inference about the phase and meter of a rhythmic stimulus by interfacing with discrete representations of its statistical regularities. The model is evaluated through computational studies which offer insights into complex and poorly understood aspects of the cognitive processing underlying rhythm perception and production: how long-term music-cultural experience shapes synchronisation to rhythmic patterns; the perception of expressive microtiming in musical rhythms; and the effect of synchronised movement on meter perception. Together, the results show that complex rhythm processing can be understood as a hierarchy of inferential predictions about a rhythmic stimulus at different temporal scales. The multiscale probabilistic models developed in this thesis are continuous-time variational Bayesian filters for point process data, with rate functions driven by variable order Markov models. These models are applicable to any time series data consisting of event sequences, and might prove particularly promising in domains such as speech research. | en_US |
