tetrak-easyocr-armenian

Armenian language support for EasyOCR — a trained recognition model, installable as a custom network.

Status: alpha, shipping v5 weights. reader() downloads a trained model and works out of the box. v5 is pre-trained on synthetic crops from eleven encyclopedia volumes and seven further works — Western Armenian literature, a scholarly history, a second encyclopedia, a bilingual dictionary — then fine-tuned on 51,078 real crops cut from 520 scanned pages. It reads 0.824 word recall on the held-out encyclopedia scans — ahead of every engine measured on those pages, including PaddleOCR's Armenian model at 0.807 and Calfa's CC BY-NC hye-calfa-n at 0.789 — and, unlike its predecessors, it holds up across registers: 0.60–0.91 word recall on eight held-out sets spanning both dialects, where v3 fell to 0.23–0.33 outside the encyclopedia it was tuned on. fold_script() now adds only ~0.005 (v5 emits Armenian forms directly rather than Latin homoglyphs), but keep it on: it never hurts. Character similarity is a separate and much weaker story, because on multi-column pages that metric measures reading order more than recognition. The numbers are on the model card.

Upgrading from v0 or v1? Do. Both were trained with 21% of their labels carrying quotation marks the images do not show, and with no class for the abbreviation dot ․ (U+2024) at all. The model card records both.

Usage

pip install tetrak-easyocr-armenian

import tetrak_hy

reader = tetrak_hy.reader()          # an easyocr.Reader, Armenian-ready
results = reader.readtext("scan.png")

reader() accepts everything easyocr.Reader does (gpu=, verbose=, …) and handles the custom-network plumbing: the network config and weights are materialised into ~/.tetrak_hy/ on first use, and the quirks of EasyOCR's custom-model loading path (there are a few) stay our problem rather than yours.

The weights live in the Hugging Face model repository, which is canonical for them, and each library version pins one immutable Hub revision — so the model you get is decided by the version of this package you installed, never by what happens to be current upstream.

Cached weights are verified against the release's SHA-256 every time they are loaded, not only when they are downloaded. A file that matches is used as it is — so a machine with no outbound access works once the weights are in place — and one that does not, because it was corrupted or because you have upgraded to a release carrying a new model, is replaced by a fresh verified download.

Set TETRAK_HY_HOME to put the cache somewhere other than your home directory, or choose it per call:

reader = tetrak_hy.reader(cache_dir="/srv/models/tetrak_hy")

A local .pth passed as weights_path= is kept in a local/ subdirectory of the cache, so testing your own weights does not disturb the released ones.

Folding cross-script homoglyphs

The recognition network has no language model, so inside an Armenian word it sometimes emits the visually identical Latin twin of an Armenian character instead — Latin h for հ, a colon for the Armenian full stop ։. fold_script() corrects these on already-recognised text, and is worth applying to every result:

results = [
    (bbox, tetrak_hy.fold_script(text), confidence)
    for bbox, text, confidence in reader.readtext("scan.png")
]

It only touches a token that already contains an Armenian letter, so Latin or Cyrillic text sharing a page is left alone. See the function's docstring for the exact scope and what it deliberately does not fold.

What it is

EasyOCR does not ship Armenian. This package adds it as a custom recognition network: EasyOCR's own generation2 architecture (VGG + BiLSTM + CTC), trained for the Armenian script — the full alphabet, the և ligature, Armenian punctuation (՝ ՛ ՞ ՜ ։ ֊ « »), digits and basic Latin for mixed material. Detection is untouched: EasyOCR's CRAFT detector already finds Armenian text; reading it is what was missing.

The import name tetrak_hy is also the EasyOCR network name — EasyOCR imports this package directly as the model architecture. If you prefer to wire the Reader yourself:

import easyocr

reader = easyocr.Reader(
    ["en"],                       # see note below
    recog_network="tetrak_hy",
    user_network_directory="~/.tetrak_hy",
    model_storage_directory="~/.tetrak_hy",
)

(["en"], not ["hy"]: EasyOCR looks up a per-language character file it does not have for Armenian. The setting is decorative for custom models — the model's own character list governs decoding — and reader() hides this entirely.)

Provenance

The model is trained by tetrak-hy-trainer on synthetic line crops: text from human-proofread pages on Armenian Wikisource (CC BY-SA 3.0) — eleven volumes of the Armenian Soviet Encyclopedia plus the collected works of Otyan, Totovents, Baronian and Tumanyan, Faustus of Byzantium, a popular medical encyclopedia and an Armenian–English dictionary — rendered in fifteen Armenian faces at real scan sizes and degraded to look scanned. v5 then adds a fine-tune on 51,078 real crops cut from 520 human-proofread scans of all twelve sources and labelled from their transcripts, mixed with the synthetic set so the model adapts to real print without forgetting the breadth it started with. The widened corpus is what taught it both dialects and several registers rather than one encyclopedia's typography; the real-crop fine-tune is what closed the gap on degraded letterpress that a cleanly rendered font cannot teach.

Every weights release carries a provenance record — data recipe, fonts, dataset revision, training config and checksums — published as provenance.json beside the weights. The training data is published too, as tetrak/armenian-ocr-crops.

Built for Tetrak, a local-first transcription pipeline for archival material, which consumes this package as its easyocr-hy backend — but nothing here depends on Tetrak.

Licence

Apache License 2.0 — see LICENSE and NOTICE. The architecture re-exported here is EasyOCR's (Apache 2.0).

Development

python3 -m venv .venv && source .venv/bin/activate
pip install -e '.[dev]'
pytest
ruff check src tests && ruff format --check src tests

Commits follow Conventional Commits, enforced by the hook in .githooks/ (git config core.hooksPath .githooks after cloning, or lefthook install). Releases and CHANGELOG.md are generated from those commits automatically on every push to main.

See CONTRIBUTING.md for the full workflow — checks, the dependency lockfile, and what the automation expects — and SECURITY.md for how to report a vulnerability.