Autoregressive Hidden Semi-Markov Model of Symbolic Music Performance for Score Following

A stochastic model of symbolic (MIDI) performance of polyphonic scores is presented and applied to score following. Stochastic modelling has been one of the most successful strategies in this field. We describe the performance as a hierarchical process of performer’s progression in the score and the production of performed notes, and represent the process as an extension of the hidden semi-Markov model. The model is compared with a previously studied model based on hidden Markov model (HMM), and reasons are given that the present model is advantageous for score following especially for scores with trills, tremolos, and arpeggios. This is also confirmed empirically by comparing the accuracy of score following and analysing the errors. We also provide a hybrid of this model and the HMM-based model which is computationally more efficient and retains the advantages of the former model. The present model yields one of the state-of-the-art score following algorithms for symbolic performance and can possibly be applicable for other music recognition problems.