Improving Rhythmic Transcriptions via Probability Models Applied Post-OMR

Despite many improvements in the recognition of graphical elements, even the best implementations of Optical Music Recognition (OMR) introduce inaccuracies in the resultant score. These errors, particularly rhythmic errors, are time consuming to fix. Most musical compositions repeat rhythms between parts and at various places throughout the score. Information about rhythmic selfsimilarity, however, has not previously been used in OMR systems. This paper describes and implements methods for using the prior probabilities for rhythmic similarities in scores produced by a commercial OMR system to correct rhythmic errors which cause a contradiction between the notes of a measure and the underlying time signature. Comparing the OMR output and post-correction results to hand-encoded scores of 37 polyphonic pieces and movements (mostly drawn from the classical repertory), the system reduces incorrect rhythms by an average of 19% (min: 2%, max: 36%). The paper includes a public release of an implementation of the model in music21 and also suggests future refinements and applications to pitch correction that could further improve the accuracy of OMR systems.