---
library_name: chant-omr
tags:
  - omr
  - gregorian-chant
  - gabc
  - openvino
  - onnx
  - square-notation
license: mit
pipeline_tag: image-to-text
---

# ChantOMR

End-to-end Optical Music Recognition for Gregorian chant square notation.
Converts photographs of historical chant manuscripts into
[GABC](https://gregorio-project.github.io/gabc/) notation.

Based on [Transcoda](https://huggingface.co/btrkeks/transcoda-59M-zeroshot-v1)'s
ConvNeXt-V2 + Transformer architecture (~59M params), retrained from scratch
on ~20,000 [GregoBase](https://gregobase.selapa.net/) scores with domain
augmentation for square notation.

## Model Details

| Property | Value |
|----------|-------|
| Architecture | ConvNeXt-V2 Tiny encoder + 8-layer Transformer decoder |
| Parameters | ~59M |
| Input | Score image (width 1050, variable height) |
| Output | GABC token sequence |
| Vocabulary | ~2048 BPE tokens |
| Training data | GregoBase (~20k scores, synthetic renders + augmentation) |
| Framework | PyTorch / Lightning |

## Evaluation

No evaluation results available yet.

## Formats

This repository contains the model in multiple formats:

| Format | Files | Use case |
|--------|-------|----------|
| **Safetensors** | `model.safetensors` | PyTorch fine-tuning or inference |
| **OpenVINO IR** | `openvino/*.xml` + `.bin` | Production inference on Intel hardware |
| **ONNX** | `onnx/*.onnx` | Portable inference on any hardware |

## Quick Start

### CLI (easiest)

```bash
pip install chant-omr
chant-omr predict score.png --model pgquiles/chant-omr --device openvino
```

### Python (OpenVINO)

```python
from chant_omr.hub import download_from_hub
from chant_omr.inference.ov_decode import load_openvino_models, ov_predict_gabc

model_dir = download_from_hub("pgquiles/chant-omr")
gabc = ov_predict_gabc(
    "score.png",
    model_dir / "openvino",
    beam_width=3,
)
print(gabc)
```

### Python (PyTorch)

```python
from chant_omr.hub import download_from_hub
from chant_omr.inference.checkpoint import load_model_from_safetensors
from chant_omr.inference.predict import predict_gabc_from_hub

model_dir = download_from_hub("pgquiles/chant-omr")
gabc = predict_gabc_from_hub("score.png", model_dir)
print(gabc)
```

## Limitations

- Trained on synthetic Gregorio renders; real manuscript accuracy depends on
  scan quality and augmentation coverage.
- Square notation only (no modern staff notation, no NABC/adiastematic neumes).
- Single-system images work best; multi-system page layout analysis is handled
  by [ghh](https://github.com/pgarciaq/ghh), not this model.

## Links

- **Code:** [github.com/pgarciaq/chant-omr](https://github.com/pgarciaq/chant-omr)
- **Pipeline:** [github.com/pgarciaq/ghh](https://github.com/pgarciaq/ghh)
- **Training data:** [GregoBase](https://gregobase.selapa.net/)

---
*Uploaded with chant-omr v0.1.0*