Hebrew Historical OCR & HTR — Custom Models, Not a Platform

Off-the-shelf engines are good enough for clean post-1850 printed Hebrew. Everything else — faded ink, Rashi, rabbinical cursive, mixed Hebrew–Aramaic–Yiddish layouts — is a research project, not a software purchase. MF Smart Research trains and operates recognition on your material, then hands back searchable text with the original page next to it.


Who this is for

  • Archives, libraries and museums sitting on Hebrew manuscripts, rabbinical print, community records or degraded newspapers that generic OCR returns as noise.
  • Universities and individual scholars who need a defensible transcription — every character traceable to a page — not a plausible paraphrase.
  • Families and independent researchers with a box of letters, a pinkas, or a run of Hebrew periodicals, who want an honest answer before committing to a full digitization.

If the collection is modern, printed, single-column Hebrew in good condition, say so on the call. Tesseract or Transkribus out of the box may be the right tool, and the honest recommendation is then to use them without us.


What a 30-minute consult looks like

This is the free scoping call described on the About page, not the pilot.

Bring (or send ahead) a few representative pages — including the difficult ones, not a cleaned sample. In thirty minutes we:

  1. Say what the pages actually are (print vs handwriting, century, damage, layout, language mix).
  2. Run the same pages through three engines and look at the output together.
  3. Share confidence scores as they come, without a global “accuracy %”.
  4. Tell you whether AI is the right tool at your scope — including “not yet,” or “a platform will do.”

If the pages look viable, the next step is a free pilot on the actual collection: for organizations, 100 documents, 10 photographs and 5 films, with no time limit; for individuals, 10 items free. Results include error rates and confidence scores. Production only starts after that.

Email [email protected] with “Consultation request” in the subject, or use the form on the English homepage.


How this differs from Transkribus

Transkribus (READ-COOP) is a professional HTR platform: cloud-based, pay-per-page, with a training UI that non-engineers can use. The 2026 Hebrew OCR guide on this site already says the quiet part: Transkribus is the right answer when you need to involve non-engineers in the training and review loop. It has a growing set of community Hebrew/Yiddish models.

MF Smart Research is not a competing HTR product.

What we do on Transkribus (and on Kraken/eScriptorium, and on custom transformer pipelines when the material demands it):

  • Decide whether Transkribus is the right environment for this collection, or whether the hard fraction needs a custom pass.
  • Layout analysis, ground-truth sampling that over-represents the difficult 20%, model training, and a review queue driven by confidence — not a random 30% check.
  • Per-collection models. Late-19th-century Vilna print, 1920s Salonika Ladino print, and an 18th-century Lithuanian rabbinical hand are three different problems.
  • Handover in standard formats, into the catalog you already use. We do not take custody of the text.

You keep the Transkribus project. We operate and train on your material. For the argument in full, see Off-the-Shelf Platform or Custom AI Pipeline?.


What the pages usually are

From work already described on this site:

  • Printed Hebrew, including 19th-century rabbinical typography.
  • Rashi script (off-the-shelf models struggle; a fine-tuned Transkribus model is the documented path).
  • Personal handwritten Hebrew letters, 20th century.
  • Rabbinical handwriting, 18th–19th century; multi-scribe volumes need per-scribe models.
  • Mixed Hebrew–Aramaic–Yiddish layouts, marginalia, bleed-through, foxing, faded ink.
  • Community ledgers — see the Hebrew page on פנקסי קהילה.

Numbers from the 2026 practitioner guide, for orientation, not as a quote for your collection:

Material Off-the-shelf After 500 lines of training After 5,000 lines
Clean modern Hebrew print 95–98% 98–99% 99%+
19th-century rabbinical print 70–82% 92–96% 97–99%
Personal handwritten Hebrew letter, 20th century 30–55% 80–90% 93–97%
Rabbinical handwriting, 18th–19th century 15–35% 70–85% 90–95%
Ladino in Rashi script 40–60% 85–93% 95–98%
Pinkas kehila with multiple hands 20–40% 65–80% 85–93%

A vendor reporting 96% without showing the test set is bluffing. We do not deliver “approximately what the document said.” We deliver what it said, with confidence, and with the original next to it.


After the text exists

OCR is the input to the next stages: language identification per region, named-entity recognition, entity linking, a searchable index. Bad OCR contaminates all of that silently. If the goal is a research instrument rather than a PDF, say so on the call.

Related reading on this site: Hebrew OCR 2026: Engines & Accuracy, When the Model Isn't Sure, Handwriting Recognition for Historical Manuscripts.


Get in touch

Request a free 30-minute consultation. Bring three difficult pages.

[email protected] · Contact form

MF Smart Research was founded by Maaty Frenkelzon, a researcher, archivist and Ph.D. candidate in History.