Improving OMR for digital music libraries with multiple recognisers and multiple sources

Large quantities of scanned music are now available in public digital music libraries. However, the information in such sources is represented as pixel data in images rather than symbolic information about the notes of a piece of music, and therefore it is opaque to musically meaningful computational processes (e.g., to search for a particular melodic pattern). Optical Music Recognition (Optical Character Recognition for music) holds out the prospect of a solution to this issue and allowing access to very large quantities of musical information in digital libraries. Despite the efforts made by the different commercial OMR developers to improve the accuracy of their systems, mistakes in the output are currently too frequent to make OMR a practical tool for bulk processing. One possibility for improving the accuracy of OMR is to use multiple recognisers and combine the results to achieve an output better than each of them individually. The general process presented here can be divided into three subtasks, S1, S2, and S3. S1 is focused in the correction of rhythmical errors at bar level, counting the errors of the different OMR outputs, establish a ranking of the results, and make a pairwise alignment to select the best measures. S2 is based on the alignment and voting of individual symbols. For this task we have implemented a conversion of the most important symbols to a simple grammar. Finally, S3 improves the output of S2 by comparing and adding symbols from S1 and detecting gaps through the alignment of wrong measures. The process described in this paper is part of our “Big Data Approach” where a large amount of data is available in music score libraries, such as the International Music Score Library Project (IMSLP), for the purpose of Music Information Retrieval (MIR).