Church Slavonic OCR

A training pipeline and a distributable Tesseract LSTM model for Church Slavonic printed liturgical texts — the accented, titlo-bearing, superscript-lettered orthography that general Cyrillic OCR models cannot read.

The model is trained on synthetic line images rendered from a large Church Slavonic text corpus in several historically-grounded typefaces, then adapted to real scans with a small hand-corrected fine-tuning set.

What's here

Path	Contents
scripts/	the Python tooling (data generation, extraction, evaluation, monitoring, review)
corpus/	the source text corpus — git submodule of slavonic/cu-md-sandbox
fonts/	the rendering typefaces — git submodules (Ponomar, Triodion, Pochaevsk, Acathist, Monomakh)
data/cu-ground-truth/	the synthetic training pairs — generated, not committed (regenerate with make dataset)
data/real-lines/	hand-corrected scan lines — committed; eval/ (held out) and finetune/
model/	the model card and evaluation reports (the cu.traineddata binary itself ships via Releases, not in the repo)
training/	tesstrain working state (checkpoints, unicharset) — not committed
docs/	the manual, reference pages, and troubleshooting

Just want to OCR?

No clone needed — download cu.traineddata from the Releases page, then:

tesseract your_line.png stdout --psm 13 -l cu --tessdata-dir /path/to/download-dir
# (or copy it into your tessdata/ dir to use -l cu directly — see model/README.md)

Train or improve the model

Start with the manual. It walks through every step, from making the practice data to publishing a release, with the exact commands and what you should see after each one.

In short:

git clone --recurse-submodules https://github.com/slavonic/church-slavonic-ocr.git
cd church-slavonic-ocr
python3 -m venv .venv && source .venv/bin/activate
make setup          # fonts, corpus, python packages (also needs Tesseract + tesstrain: manual, Step 1)
make dataset        # Step 2: synthetic training lines
make train-seeded   # Step 3: base model
make review-staging # Step 4: correct real lines (after extract_lines.py)
make eval           # Step 5: score on held-out real lines

Results

Download the trained model from the Releases page. See the model card for current metrics and model/eval/ for the per-line report. Character error rate is reported two ways — full, and with combining marks stripped — so letterform accuracy is separable from diacritic placement (see evaluation).

Documentation

Start here: docs/manual.md, the step-by-step guide from nothing to a published model.

Reference, for when you need the details:

• docs/data-generation.md — cu_make_training_data.py: fonts, charset, hyphenation, degradation
• docs/training.md — seeded vs from-scratch training, the fine-tune rules, resetting the charset
• docs/evaluation.md — every extract_lines.py option, and how to read cu_eval.py's output
• docs/troubleshooting.md — the failure modes we actually hit, grouped by step
• data/README.md — what's in data/, and ground-truth typing conventions

License & attribution

This repository combines works under different licenses; please respect each:

• Code (scripts/) — see LICENSE (MIT License).
• Corpus (corpus/, submodule) — retains the license of slavonic/cu-md-sandbox (MIT License).
• Fonts (fonts/, submodules) — each retains its own license (SIL Open Font License).
• Model (model/cu.traineddata) — a derivative work of the corpus and fonts; its provenance and intended use are documented in the model card.