An Efficient State-Space Model for Joint Tempo and Meter Tracking

Dynamic Bayesian networks (e.g., Hidden Markov Models) are popular frameworks for meter tracking in music because they are able to incorporate prior knowledge about the dynamics of rhythmic parameters (tempo, meter, rhythmic patterns, etc.). One popular example is the bar pointer model, which enables joint inference of these rhythmic parameters from a piece of music. While this allows the mutual dependencies between these parameters to be exploited, it also increases the computational complexity of the models. In this paper, we propose a new state-space discretisation and tempo transition model for this class of models that can act as a drop-in replacement and not only increases the beat and downbeat tracking accuracy, but also reduces time and memory complexity drastically. We incorporate the new model into two state-of-the-art beat and meter tracking systems, and demonstrate its superiority to the original models on six datasets.