Optical Music Recognition

Optical Music Recognition (OMR) turns score images into symbolic formats such as MusicXML, ABC, or **kern, enabling digitization, search, and computational analysis of music. Approaches range from convolutional and recurrent sequence models trained with CTC to vision transformers, autoregressive vision-language models, and grammar-guided decoding; synthetic data and pretraining are common ways to address limited annotations. Evaluation uses datasets of typeset, scanned, handwritten, historical, and chant notation, with measures such as symbol error rate and OMR-NED that assess notation-specific mistakes. A central challenge is generalizing across layouts, instruments, notation systems, and manuscript styles, while benchmark results can also depend strongly on encoding, segmentation, and score-level data splits.