| --- |
| license: mit |
| task_categories: [image-to-text] |
| tags: [optical-music-recognition, omr, music, jianpu, sargam, synthetic] |
| size_categories: [10K<n<100K] |
| --- |
| |
| # omr-synth — 30,000 synthetic music-notation pages (Western · Jianpu · Sargam) |
|
|
| Render-forward synthetic data for optical music recognition of the notations general vision models can't |
| read. **One symbolic score → three renderers → one shared OMRDSL target.** Every image ships with a |
| *perfect* label — no human transcription. Built for a weekend learning project. |
|
|
| - 💻 Code: https://github.com/abhid1234/omr-synth |
| - 🤖 Model: https://huggingface.co/abhid1234/omr-synth |
| - ▶️ Demo: https://omr-synth.vercel.app |
|
|
| ## Contents |
| - **`omr-synth-30k.tar.gz`** — 30,000 `(image, target)` pairs: `images/*.png` + `targets/*.omrdsl`. |
| 10,000 each of **Western staff**, **Jianpu (简谱)**, **Sargam**. Curriculum levels 0–3 (clean → manuscript-like). |
| - **`manifest.jsonl`** — one row per example: `id, image, target, notation, renderer, curriculum, width, height`. |
| - **`samples/`** — a small browsable preview. |
|
|
| ## How it's made |
| Western: MusicXML → Verovio → CairoSVG. Jianpu & Sargam: drawn programmatically from the *same* symbolic |
| record. Deterministic by seed; fully regenerable from the repo (`make synth SIZE=n`). |
|
|
| ## Splits |
| Stable train/val/test (80/10/10) by hashing each `id` (`src/data/split.py`). MIT · Abhi Das. |
|
|
