Improving Optical Music Recognition by Combining Outputs from Multiple Sources

Current software for Optical Music Recognition (OMR) produces outputs with too many errors that render it an unrealistic option for the production of a large corpus of symbolic music files. In this paper, we propose a system which applies image pre-processing techniques to scans of scores and combines the outputs of different commercial OMR programs when applied to images of different scores of the same piece of music. As a result of this procedure, the combined output has around 50% fewer errors when compared to the output of any one OMR program. Image pre-processing splits scores into separate movements and sections and removes ossia staves which confuse OMR software. Post-processing aligns the outputs from different OMR programs and from different sources, rejecting outputs with the most errors and using majority voting to determine the likely correct details. Our software produces output in MusicXML, concentrating on accurate pitch and rhythm and ignoring grace notes. Results of tests on the six string quartets by Mozart dedicated to Joseph Haydn and the first six piano sonatas by Mozart are presented, showing an average recognition rate of around 95%.