Optical music recognition for medieval chant has been held back by a simple problem: there are very few datasets that connect manuscript images to the musical symbols they contain. Without such ground truth, systems cannot learn to read the diverse notational styles used in chant manuscripts.

To fill this gap, the authors turned to the Cantus Database, which already holds thousands of Volpiano transcriptions—a text-based encoding of chant melodies. By aligning those existing transcriptions with corresponding manuscript images, they produced 30,567 lines of ground truth.

The resulting dataset spans different notational families, making it broadly useful for training and evaluating OMR models. Because the method leverages already-curated transcriptions, it also demonstrates a practical way to expand ground truth without starting from scratch.