18 December 2002, Music & AI Colloquium, Taylan Cemgil

Automatic music transcription refers to extraction of a human readable and interpretable description from a recording of a musical performance. Traditional music notation is such a description that lists the pitch levels (notes) and corresponding timestamps. Such a representation would be useful in several applications such as interactive music performance, information retrieval (Music-IR) and content description of musical material in large music databases. In this talk, I will focus on a subproblem in music-ir, where I assume that exact timing information of notes is available, for example as a stream of MIDI events from a digital keyboard. I will present a probabilistic generative model for timing deviations in expressive music performance. The structure of the proposed model will turn out to be a switching state space model (switching Kalman filter). The switch variables correspond to discrete note locations as in a musical score. The continuous hidden variables denote the tempo.

Given the model, we can formulate two well known music recognition problems, namely tempo tracking and automatic transcription (rhythm quantization) as filtering and maximum a posteriori (MAP) state estimation tasks. Unfortunately, exact computation of posterior features such as the MAP state is intractable in this model class, so we resort to Monte Carlo methods for integration and optimization. I have compared Markov Chain Monte Carlo (MCMC) methods (such as Gibbs sampling, simulated annealing and iterative improvement) and sequential Monte Carlo methods (particle filters). Simulation results suggest better results with sequential methods. The methods can be applied in both online and batch scenarios (such as tempo tracking and transcription) and are thus potentially useful in a number of music applications such as adaptive automatic accompaniment, score typesetting and music information retrieval.