Synthesised data is a common strategy to address the lack of annotated handwritten music scores in Optical Music Recognition (OMR). Smashcima is a recent synthesis framework that assembles pages of handwritten scores by sampling glyph images from MUSCIMA++ and positioning them with hard-coded rules. In this work we ask whether these rules are grounded empirically and whether a learned model for context-dependent positioning can further improve synthesis quality. We extract inter-glyph distributions from around 285 thousand annotated relationships across MUSCIMA++ and OmniOMR, finding that several key parameters deviate systematically from real handwriting. Replacing these constants with empirical means reduces positioning RMSE by up to 60%. We then train a Graph Neural Network (GNN) on the Music Notation Graph structure to predict context-dependent offsets, achieving an additional 50 to 85% RMSE reduction on stems and beams by modelling how neighbouring glyphs constrain placement. Cross-dataset analysis shows that spacing differences between the two corpora are explained by note density (3.3), not style, and that duration-proportional spacing does not hold in handwriting (R2 = 0.001). In downstream evaluation, a YOLOv8 detector pretrained on GNN-positioned synthetic data achieves +30% mAP50 over the baseline and +11% over empirical means when fine-tuned on 200 real handwritten staves, showing that positioning accuracy matters for downstream OMR.
Figures & tables
Figure 1: Empirical positioning distributions. Row 1: cross-dataset scatterplots (blue: MUSCIMA++, orange: OmniOMR). Row 2: key findings with Smashcima baselines (red dashed) and empirical means (green). Row 3: writer-dependent beam style, inter-note spacing, and the absence of duration-proportional spacing ( R2≈0.001 ).
Relationship
Baseline
Empirical
RMSE
Improv.
value
mean
(global)
Stem length (up)
5.0mm
3.8mm
1.44mm
+60%
Stem length (down)
5.0mm
4.2mm
1.32mm
+57%
Beam angle (global)
0\SIUnitSymbolDegree
4.0\SIUnitSymbolDegree
4.51\SIUnitSymbolDegree
+40%
Beam angle (per-writer)
0\SIUnitSymbolDegree
writer int.
3.81\SIUnitSymbolDegree
+49%
Inter-note gap
1.2mm
5.07mm
3.94mm
+29%
Table 1: Summary of positioning corrections. RMSE improvement is relative to Smashcima’s hard-coded baseline, evaluated under LOWO cross-validation.
Edge type
Baseline
Empirical
GNN
vs Emp.
n
notehead → stem
3.07
1.44
0.60
+80.6%
1984
notehead → beam
6.89
7.13
1.07
+84.4%
1743
notehead → flag
3.68
3.68
0.75
+79.5%
323
notehead → ledger line
1.16
1.16
0.55
+52.7%
600
notehead → staccato
2.59
2.59
0.91
+65.0%
289
notehead → duration dot
0.62
0.62
0.61
+1.5%
131
Table 2: Per-edge-type positioning RMSE (mm). The GNN improves most on relationships where placement is constrained by local context.
Figure 2: LOWO cross-validation: per-writer RMSE distribution across 50 held-out writers. The GNN generalises consistently (mean 2.06±0.30 mm), with only two outliers above 3 mm.
Figure 3: Synthesised output from the same MusicXML input and random seed. Left: baseline (hard-coded constants). Centre: empirical (data-driven means). Right: GNN (context-dependent positioning). Note the progressively shorter stems, tilted beams, and wider inter-note spacing from left to right. The GNN output shows more realistic stem length variation in beamed groups.
Shots
Baseline
Empirical
GNN
Δ vs Base
Δ vs Emp
5
0.093
0.101
0.114
+22.6%
+12.9%
10
0.104
0.106
0.125
+20.8%
+18.3%
20
0.110
0.109
0.126
+15.1%
+16.1%
50
0.115
0.116
0.136
+18.7%
+17.5%
100
0.123
0.137
0.153
+24.0%
+11.4%
200
0.168
0.196
0.218
+30.0%
+11.3%
Table 3: Few-shot fine-tuning on real handwritten staves (mAP50). The GNN-pretrained model consistently outperforms both baselines across all data regimes.
Figure 4: Few-shot domain adaptation on real OmniOMR staves. The GNN-pretrained model (green) consistently outperforms both the empirical (orange) and baseline (red) models across all data regimes. The advantage is largest in the low-data regime where pretraining quality matters most.
Optical Music Recognition (OMR) has seen major progress in model design, with end-to-end methods now capable of recognising notation at all levels of complexity. However, the impact of this progress has been limited by the visual domains of available training datasets, which are largely born-digital. Existing large collections of sheet music in libraries and other heritage institutions contain predominantly manuscripts, whose visual domains are highly diverse and different, so existing OMR systems fail when applied in the real world. These institutions are often resource-constrained, so large in-domain datasets cannot be expected. We provide a first baseline on real-world manuscripts with complex piano notation in the resource-constrained scenario. Using fine-grained music notation graph (MuNG) annotations and the Smashcima synthesis tool, we then show that while some direct transcriptions of in-domain data remain essential, domain adaptation using synthetic musical manuscript images brings significant improvement. Furthermore, the symbols used do not need to be in-domain, so the expensive fine-grained annotation can be avoided. We thus bring OMR closer to one of its stated goals: preserving and promoting musical cultural heritage.
Jiří Mayer, Martina Dvořáková, Vojtěch Dvořák +6
Institute of Formal and Applied Linguistics, Charles University · Moravian Library
Optical Music Recognition (OMR), the task of transcribing sheet music into a structured textual representation, is currently bottlenecked by a lack of large-scale, annotated datasets of real scans. This forces models to rely on either few-shot transfer or synthetic training pipelines that remain overly simplistic. A secondary challenge is encoding non-uniqueness: in the popular Humdrum **kern format for transcribing music, multiple different text encodings can render into the same visual sheet music. This one-to-many mapping creates a harder learning task and introduces high uncertainty during decoding. We propose Transcoda, an OMR system built on (i) an advanced synthetic data generation pipeline, (ii) a normalization of the **kern encoding to enforce a unique normal form and (iii) grammar-based decoding to ensure the syntactic correctness of the output. This approach allows us to train a compact 59M-parameter model in just 6 hours on a single GPU that outperforms billion-parameter baselines. Transcoda achieves the best score among state of the art baselines on a newly curated benchmark of synthetically rendered scores at 18.46% OMR-NED (compared to 43.91% for the next-best system, Legato) and reduces the error rate on historical Polish scans to 63.97% OMR-NED (down from 80.16% for SMT++).
Full-page end-to-end Optical Music Recognition seeks to transcribe entire music pages directly into symbolic notation, avoiding the limitations of traditional pipelines that rely on accurate staff segmentation. Recent Transformer-based architectures have achieved strong performance on typeset scores, relying on large-scale synthetic data for pretraining. However, their applicability to handwritten music remains largely unexplored. In this work, we study full-page transcription on handwritten monophonic collections and analyze the impact of synthetic pretraining in this setting. To investigate which factors are most relevant during pretraining, we introduce a generator capable of producing visually coherent full-page scores in both typeset and handwritten styles. Experiments on three real handwritten datasets provide a comparative evaluation of several full-page pipelines and different synthetic pretraining strategies. The results suggest that the benefits of synthetic pretraining are primarily associated with learning structural layout conventions rather than with visual similarity to the target handwriting.
Adrian Rosello, Antonio Ríos-Vila, David Rizo +1
Pattern Recognition and Artificial Intelligence group, University of Alicante, Alicante, Spain · Instituto Superior de Enseñanzas Artísticas de la Comunidad Valenciana, Alicante, Spain