Yiddish OCR & Handwriting Recognition

“Yiddish OCR” is not a single switch. Mid-20th-century printed Yiddish, Vaybertaytsh (the semi-cursive Ashkenazi type used across early modern Europe), and late-19th / early-20th-century handwriting are three recognition tasks. Off-the-shelf Hebrew engines often pick the wrong language on mixed pages and corrupt the rest. MF Smart Research runs the model that matches the page — including public community models on Transkribus — and trains on your material when those models were built on someone else’s.


Who this is for

  • Archives and libraries with Yiddish periodicals, correspondence, personal documents, or community regulations.
  • Scholars working through a corpus of letters, play manuscripts, or takanot who need searchable text rather than a folder of JPEGs.
  • Families holding Yiddish letters or a pinkas written in Yiddish or mixed Hebrew–Yiddish.

Selected work already listed on the About page includes OCR and HTR pipelines for early-20th-century Yiddish printed and handwritten materials — periodicals, correspondence and personal documents. No institution names are published there; none are invented here.


What a 30-minute consult looks like

Free scoping call, as on the About page.

Send a few pages that are actually typical: a dense newspaper column, a letter in a tight hand, a Vaybertaytsh title page, a ledger line — not only the cleanest scan. In thirty minutes we:

  1. Classify the script (modern print, older mashket/Vaybertaytsh, handwriting, mixed Hebrew).
  2. Run the same pages through three engines. Where a public Transkribus model is the obvious first try, that is one of the three.
  3. Read the errors together (systematic letter confusions vs. layout collapse vs. language ID).
  4. Say whether to stay inside Transkribus, fine-tune on your pages, or move the hard fraction to a custom pipeline — including “not yet.”

A viable case goes to 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 — with error rates and confidence scores.


Community models we can operate (not ours)

Two public Transkribus models are the ones people mean when they say “there is already Yiddish HTR”:

  • The Dybbuk for Yiddish Handwriting (Transkribus model ID 46159; published on the Transkribus model hub, 13 November 2022). Ground truth described by the publishers: mainly play manuscripts from the late 19th and early 20th century, written in New York and Europe, plus letters corrected in a public workshop. Built by the DYBBUK project team (see the Transkribus model card and In geveb, Feb 2022). MF Smart Research did not publish this model.
  • Vaybertaytsh.YidTakNL (Transkribus model ID 57147; Ronny Reshef and Mirjam Gutschow, 29 November 2023). Trained on 18th–19th-century Amsterdam community regulations in Vaybertaytsh type; built to support research at Erasmus University. The publishers note it is based on an earlier model which itself is based on Dybbuk. See the model card and the JOHD data paper. Not an MF-SR model.

Published validation CER on a model card is measured on that model’s own validation set. Transkribus says so on the card. It is not a quote for your letters, your pinkas, or your newspaper.

What we do: run the relevant public model on a sample from your collection, measure it against a small ground-truth set you approve, and only then decide whether to fine-tune, switch engines, or stay with manual correction. Below roughly 500 pages, the 2026 guide’s rule still holds: manual correction is usually cheaper than training.

Other public Hebrew-script models on the same hub (for example DiJeSt for printed Hebrew-script material) are in the same category: tools we can operate, not assets we claim.


How this differs from Transkribus

Transkribus is the platform. For Yiddish it is often the correct place to work, because the community models live there and because non-engineers can correct lines in the same UI that trains the next model.

MF Smart Research does not replace that platform. We operate it: layout, sampling, training, confidence triage, and export into your catalog. When the page is not what Dybbuk or Vaybertaytsh.YidTakNL saw in training — a different hand, a damaged pinkas, Yiddish mixed with Russian stamps — we say so, and either fine-tune on your pages or move that fraction to a custom pipeline (Kraken, or a transformer stack). That is the hybrid described in Off-the-Shelf Platform or Custom AI Pipeline?.


Numbers already on this site

From the 2026 practitioner guide:

Material Off-the-shelf After 500 lines After 5,000 lines
Yiddish print, mid-20th century 80–90% 95–98% 98–99%
Pinkas kehila with multiple hands 20–40% 65–80% 85–93%

Train a custom model when (a) accuracy on a representative sample is below 90%, (b) the document type has distinctive features such as Yiddish–Hebrew mixing or worn ink, or (c) the project exceeds 500 pages.


Get in touch

Request a free 30-minute consultation. Three typical pages are enough to start.

[email protected] · Contact form

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

Longer treatment of the same ground: Yiddish Handwriting OCR in 2026.