Tetrak OCR
Turns scans of unruly archive material into searchable text, on your own machine, choosing the best engine for each file and setting aside what it cannot read.
Historical archives do not arrive as tidy, uniform PDFs. A real collection is a chaotic mix of linen postcards, dense multi-column trade journals, illustrated theatre handbills and hand-lettered vaudeville posters. One-size-fits-all OCR fails on this material because it assumes clean, high-contrast, linearly structured print.
Tetrak OCR is the first of Tetrak’s tools: a small, local-first program that turns those scans into full-text transcripts — text you can attach to the images in your catalogue or digital asset management system, so that collections become searchable by their contents, hard-to-read items gain a readable transcript, and screen readers have a text alternative to work with. It inspects each document, runs the viable local engines over it, scores their output without a reference transcript, and sets aside anything too poor to trust.
auto-local fans out over the engines eligible for that file — four at most, two for a PDF. Vision is macOS-only and is not currently in that pool, though it measures as the strongest local engine on the corpus.One interface, multiple engines
Tesseract, EasyOCR, PaddleOCR, Marker, Apple Vision and Claude behind a single call. Swap engines with a flag; nothing else in your pipeline changes.
auto-localpicks for youRuns every viable local engine on each file, scores the results with no reference text, and keeps the best one. Per file, not per batch. Name a single engine instead when throughput matters more.
Everything stays on your machine
No cloud account, no API key, no material leaving the building. The one cloud path — a vision-model escalation for pages nothing local can read — is opt-in, per file, invoked by a person.
It tells you when it failed
Anything below the quality floor goes to a
workspace/triage/queue with a manifest of what each engine produced. Bad transcripts do not enter your archive looking like good ones.
Two problems, one tool
No single OCR engine wins on mixed material. Measured over a corpus of Los Angeles stage and picture-palace ephemera (c. 1910–1945), five different engines win across nine documents. There is no setting you can choose once and be right:
| Your archival material | Best local engine | Character similarity | Word recall |
|---|---|---|---|
| Printed captions and postcard reverses | tesseract-auto | 0.91 | 0.71 |
| Low-contrast, textured illustration fronts | easyocr | 0.87 | 0.71 |
| Illustrated newspaper advertising | marker | 0.75 | 0.88 |
| Multi-page souvenir programmes | marker | 0.66 | 0.93 |
| Dense small type on a playbill | paddle | 0.39 | 0.90 |
| Hand-lettered display type | paddle | 0.13 | 0.39 |
| Mixed collections, unattended | auto-local | 0.57 | 0.76 |
Choosing per file rather than per batch is worth a real margin: auto-local
beats any single local engine on both metrics, entirely offline and free.
The research walks through how that was measured, and
the detail has the full benchmark, the method and its
honest caveats.
Some material cannot go to a cloud service. Donated collections often arrive with promises attached — donor agreements, privacy constraints, material about living people — and many institutions simply do not want collection scans leaving the building, least of all into services that might train on them. Tetrak OCR’s answer is structural rather than contractual: processing happens on your machine, nothing requires an account or a key, and the one path that sends a file to an API exists only as a deliberate, per-file decision by a person looking at the triage queue.
Who it is for
Curators of special collections, regional and county archives, local history societies, community archives, and museum collections rich in ephemera — the institutions with the most unruly material and the least appetite for cloud contracts. It runs on ordinary hardware; the reference environment is a laptop. If your aim is publishing out-of-copyright material openly, the full pipeline applies too, escalation included; if your collections carry restrictions, everything useful happens without any material leaving your machine.
Get it
pip install "tetrak[all]"
tetrak-ocr batch --backend auto-local
The tutorial is the three-minute version; the user guide covers installation, the optional extras, the command line and the Python API in full.
What it does not do yet
- Handwriting. Every engine it routes between reads print. A handwritten letter or a manuscript page goes to the triage queue, where a person decides.
- A reference-free score is a heuristic, not a measurement. The quality gate ranks what the engines produced; it cannot know what the page says. The research sets out how well that ranking agrees with human-checked transcripts, and where it does not.
- Every engine reads every format. PDFs reach Tesseract, Marker and the cloud escalation; EasyOCR and PaddleOCR take images only, and the batch pipeline skips the files an engine cannot open rather than failing on them.
- Scripts beyond Latin and Armenian. Tesseract and PaddleOCR bring their own language packs; the recogniser we train ourselves is Armenian so far.
Use it with your catalogue
A transcript is only useful once it sits beside the image in the system your collection already lives in. tetrak-omeka is a working demonstration with Omeka S: a local instance, scripts that send its images through Tetrak OCR and put the transcripts back, and tutorials that walk through doing the same with your own material, with or without any code.
Articles about it
- What we will be writing about: why the tool exists and what the site measures.
- The engine that made it up: two compact vision-language models tried as local engines, and the one that invented words.
- Which way up is the page?: working out which scanned pages are on their side before reading them.
Where to go next
Install it, transcribe one image and batch a small folder, in three steps.
Installation and the optional extras, every command and flag, the Python API, and tutorials that go further.
Nine documents, five engines and what each one breaks, in five beats — then the corpus, the method, the full benchmark, which tool to reach for, and how each finding became a mechanism, including what the numbers do not establish, and one retracted finding.
The corpus, the engines as measured, routing and licensing, and where the user guide documents the command line and the Python API.