Work

Mixing Real and Synthetic Data for Handwritten Optical Music Recognition: Domain Adaptation on Debussy Manuscripts

Year: 2026

Type: conference-paper

Abstract: We fine-tune a Sheet Music Transformer on our dataset of Claude Debussy's autograph manuscripts [10], cutting handwritten optical music recognition (OMR) symbol error rate (SER) from 84% zero-shot to 51% fine-tuned. Mixing in synthetic data (via Smashcima [3]) improves SER by only 4 more points: deterministic generators don't yet cover notation missing from pretraining, such as slurs and articulations, which account for 1.9× their share of errors. OMR benchmarks are almost entirely synthetic, and handwritten score datasets remain scarce. The Debussy dataset [10] pairs Claude Debussy's autograph piano manuscripts, held at the BnF, with MusicXML ground truth at the page and system levels.

Source: HAL

Authors: Scheithauer, Hugo, Clérice, Thibault, Romero-García, Gonzalo

Institutions: ALMANACH: Modélisation et analyse linguistique automatique et humanités computationnelles, École Pour l'Informatique et les Techniques Avancées

Language: English

Topic: Handwritten Text Recognition Techniques

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

OpenAlex ID: W7213809336

FWCI:

Cites:

Cited by:

Related to: 10