How to be lost: Principled Priming and Pruning with particles in Score following

Previous work in score following has provided methods for aligning a skilled live performance to a symbolic or audio score. In the Bayesian framework, ideal generative models require O(n) computations at each real time step where n is the length of the score. In practice, heuristic thresholds have been used to consider only a subspace of generative models with high priors conditioned on the previous state. These heuristics work well for skilled performances but fail when large errors are made by amateur musicians. We present a novel Priming Particle Filter for audio scores which places the order-limiting heuristic on a firm foundation and adds the ability to recover from large errors by using psychologically-inspired bottom-up priming in addition to regular sequential importance sampling.