Armenian text recognition

A recogniser we trained ourselves for printed Armenian, published as a small library that adds the language to EasyOCR, with the model, the word list and the training data in the open.

What it is for

Most OCR (optical character recognition) software reads Armenian badly or not at all. The free engines either have no Armenian model, or have one trained on so little that a page of a 1950s encyclopedia or a diaspora newspaper comes back half wrong. A collection with Armenian books, pamphlets and periodicals in it cannot be searched by its contents until something can read them.

This recogniser is our answer. It reads printed Armenian well enough to make a scanned page findable, and it is free to install, free to fork and free to retrain.

Who it is for

Libraries, archives and community collections with Armenian-language holdings; anyone building a digitisation workflow that has to handle Armenian; and researchers who want a recogniser they can measure, retrain or extend. If you already use EasyOCR, it is one pip install away. If you use Tetrak OCR, it is already the engine behind the easyocr-hy backend.

What it does today

• Reads printed Armenian across the full alphabet, the և ligature, Armenian punctuation, digits and the Latin letters that mixed pages carry. Detection, finding the lines on the page, is EasyOCR’s own; reading them is what was missing.
• Both the Eastern and the Western language, in either spelling. The training corpus runs from the Armenian Soviet Encyclopedia to literary editions in both dialects, so it has seen the reformed orthography used in Armenia and the classical one the diaspora kept.
• Corrects itself with a word list. An optional list of 1.1 million Armenian word forms lets the model fix a word it misread by one letter, when the right word was its own second choice.
• Folds look-alike letters. The model sometimes writes a Latin letter for its Armenian twin; a helper puts the Armenian back without touching the Latin text on the same page.
• Ships with its provenance. Every release carries a record of the data, fonts, training settings and checksums that produced it, and the training data itself is published.

How well it reads is measured on held-out pages from eight kinds of printed Armenian, against every other Armenian engine we can run. The numbers, and the kinds of page it still trails on, are in the articles below and on the model card; we do not copy them into prose, where they would go stale.

What it does not do yet

• Handwriting. It reads print. Manuscripts and handwritten letters are a different model, not yet trained.
• Layout. It reads lines; it does not reconstruct columns, tables or reading order. Tetrak OCR handles the ordering when the recogniser runs inside it.
• Other scripts. Georgian and Syriac recognisers are where we would like to go next; they do not exist yet. Our transliteration library already handles those scripts for search.
• Perfect pages. It is alpha software and some registers of print still defeat it. The model card says which.

Get it

pip install tetrak-easyocr-armenian

import tetrak_hy

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

Inside Tetrak OCR, pip install "tetrak[armenian]" installs the same model as the easyocr-hy backend. The weights download on first use and are checked against the release’s checksum every time they load.

Articles about it

The story of the model, newest first, with the measurements that go with each step.

• Beating the baselines, starting with our own ruler: how v6 came to lead the other engines, and the metric mistake found on the way.
• When your training set has read your test set: keeping the evaluation honest.
• Telling classical from reformed Armenian by counting one letter: the two spellings, and how to tell them apart.
• The engine we had left out and The comparison we had been putting off: every Armenian engine we can run, measured on the same pages.
• Built to be forked: how to pick up the trainer and the library.
• The Armenian recogniser learns both dialects and Shipping the Armenian recogniser: the first versions, and what the language taught the model.
• Armenian OCR: the engines, the numbers, and our first training run: where it started.