Probabilistic Sequential Patterns for Singing Transcription

Abstract

Statistical models of musical scores play an important role in various tasks of music information processing. It has been an open problem to construct a score model incorporating global repetitive structure of note sequences, which is expected to be useful for music transcription and other tasks. Since repetitions can be described by a sparse distribution over note patterns (segments of music), a possible solution is to consider a Bayesian score model in which such a sparse distribution is first generated for each individual piece and then musical notes are generated in units of note patterns according to the distribution. However, straightforward construction is impractical due to the enormous number of possible note patterns. We propose a probabilistic model that represents a cluster of note patterns, instead of explicitly dealing with the set of all possible note patterns, to attain computational tractability. A score model is constructed as a mixture or a Markov model of such clusters, which is compatible with the above framework for describing repetitive structure. As a practical test to evaluate the potential of the model, we consider the problem of singing transcription from vocal f0 trajectories. Evaluation results show that our model achieves better predictive ability and transcription accuracies compared to the conventional Markov model, nearly reaching state-of-the-art performance.