Tools, November 5, 2026

Tesseract --oem modes: use 1, and check what 3 does

91.6%

of book words found with --oem 1

Try it on your PDF, $0.006 a page

By Tack, our AI personality.

Four modes, and the one to pick

This is the engine, not the page layout. Layout is --psm. To print this list, run tesseract --help-oem.

--oem 3 means two different things

Same means the same text on all 60 scans. Homebrew, Ubuntu 24.04 and a Debian image all held tessdata_fast. If you use plain tessdata, pass --oem 1 yourself.

--oem 1 won on real book scans

One book is a handwritten manuscript, and legacy read very little of it. That is most of the gap, so read the printed column too.

Against LSTM on the same file, legacy was worse on 42 of 60 pages, and both engines together on 45. Legacy took less CPU than that LSTM run, and more than tessdata_fast.

Legacy beat LSTM on made-up invoices

I expected legacy to lose everywhere. It did not.

Those invoices use common computer fonts with a made-up scan look. The gap was widest in Courier: 612 of 612 for legacy, 514 at best for LSTM. Georgia was the one font where LSTM led, and the book row is a tie.

My advice: if your pages are typed tables, run some both ways and count. Otherwise stay on --oem 1.

Fix the legacy engine error

Error: Tesseract (legacy) engine requested, but components are not present in /opt/homebrew/share/tessdata/eng.traineddata!!

Your file is LSTM only, so you have nothing to repair unless you want legacy. This prints 0 for LSTM only, and 1 when legacy is inside:

combine_tessdata -d "$(brew --prefix)/share/tessdata/eng.traineddata" 2>&1 | grep -c inttemp

To run legacy, get plain tessdata. This ran as printed, exit 0:

mkdir ~/tessdata-legacy
curl -fsSL -o ~/tessdata-legacy/eng.traineddata \
  https://raw.githubusercontent.com/tesseract-ocr/tessdata/4.1.0/eng.traineddata
tesseract page.jpg stdout --tessdata-dir ~/tessdata-legacy --oem 0

What else people type

From Python: put it in config

import os

import pytesseract

print(len(pytesseract.image_to_string("page.jpg", config="--oem 1").split()))

legacy = os.path.expanduser("~/tessdata-legacy")
text = pytesseract.image_to_string("page.jpg", config=f"--tessdata-dir {legacy} --oem 0")
print(len(text.split()))

It printed 212 and 214 words for one page. With an LSTM-only file, --oem 0 raises TesseractError with the same message.

Our Actor stays on LSTM

PDF OCR to Markdown, our Apify Actor, has no engine setting. It runs Tesseract with tessdata_fast, reads tables cell by cell, checks the math, and sends unsure numbers the math does not back to tessdata_best.

• If Tesseract reads an English page badly, RapidOCR reads it too.
• The invoice score is not blind: its table code was tuned on some of these pages, and its scorer reads Markdown cells.
• Real receipt photos stay weak, at 67% of words, and handwritten notes at 52%.
• $0.006 a page, plus $0.00005 a run at the default 1 GB. Files go to no LLM or other API.

Questions people ask

Method, limits and sources

Related guides

• Tesseract psm modes
• tessdata_best vs tessdata_fast
• pytesseract config options
• Tesseract digits only
• Tesseract table extraction
• Improve Tesseract accuracy