Mon OCR

Optical Character Recognition for Mon (mnw) text.

Mon is classified as vulnerable in UNESCO's Atlas of the World's Languages in Danger. General OCR covers its script but not its language, and the difference is the reason this package exists.

Mon is written in the Mon–Burmese script, so a Burmese recogniser will read a Mon page and return text. Google Cloud Vision lists Burmese (my, Mymr) as an experimental OCR language and has no entry for Mon; Cloud Translation and Google Translate's 2024 expansion both ship Burmese and not Mon; Tesseract's Myanmar.traineddata is a script model, which by design covers languages it was never trained on.

What that misses is specific and checkable. Ten of the 276 characters this model emits are named for Mon in Unicode precisely because Burmese does not use them: U+1028 MON E, U+1033 MON II, U+1034 MON O, and U+105A–U+1060 (MON NGA, JHA, BBA, BBE, and the MON MEDIAL NA, MA and LA signs). A Burmese-trained model has no output class for any of them, so it emits the nearest Burmese glyph instead. The result is fluent-looking Burmese-Mon hybrid text with nothing raised — wrong in a way that is hard to notice unless you read Mon.

This package is a Mon recogniser you can pip install, running locally on ONNX Runtime. It reads printed Mon: rendered pages and scans of them. Handwriting is out of scope. The held-out result and, more importantly, the four things it does not cover are on the model card.

Installation

pip install monocr

or, in a uv project:

uv add monocr

Quick Start

Python Usage

from monocr import MonOCR

# Initialize. Downloads the model pinned in config.HF_REVISION on first use.
model = MonOCR()

# A page: segmented into lines, joined with newlines
print(model.predict_page("page.png"))

# A line crop: read whole, never split
print(model.predict_line("line.png"))

# With a confidence score
result = model.predict_with_confidence("page.png")
print(f"Text: {result['text']}")
print(f"Confidence: {result['confidence']:.2%}")

predict and read_text are aliases for predict_page. If you are passing single-line crops, call predict_line — it is the only path that cannot split one line into several.

Examples

See the examples/ folder to learn more.

• examples/run_ocr.py: A complete script that can process a folder of images or read a full PDF book.
• Or a demo notebook to play around with the package notebooks/demo.ipynb

CLI Usage

You can also use the command line interface:

# Process a single image
monocr read image.png

# Process a folder of images
monocr batch folder/path

# Manually download the model
monocr download

Resources

• monocr on pypi
• monocr on hugging face
• Changelog

Development

uv sync
uv run pytest

Release Workflow

pyproject.toml holds the only version in the tree; monocr.__version__ reads it back from the installed metadata. The tag has to match it, and the release workflow fails the build if it does not.

uv version --bump patch          # or minor / major
git commit -am "release: $(uv version --short)"
git tag "v$(uv version --short)"
git push origin HEAD --tags

Two things this section got wrong before, both caught by running it:

• It showed git tag v2.2.3 beside a version of 2.2.0. Nothing would have stopped that going out; release.yml now compares them.
• These four commands alone used to leave a release broken. __version__ was a separate literal in __init__.py, uv version --bump does not touch it, and the test that noticed ran after the tag was already pushed. The version is derived now, so the commands above are the whole procedure.

Publishing runs from the tag, through GitHub Actions, using PyPI trusted publishing. It installs from the lockfile and runs the test suite first.

License

MIT - do whatever you want with it.