# SOTA OCR (state of the art optical character recognition)

**URL:** <https://forum.biodiversitylibrary.org/t/sota-ocr-state-of-the-art-optical-character-recognition/42>\
**Category:** Technology and Tools\
**Tags:** ai, ocr\
**Created:** [March 25, 2026, 4:29pm UTC](https://forum.biodiversitylibrary.org/t/sota-ocr-state-of-the-art-optical-character-recognition/42 "2026-03-25T16:29:40Z")\
**Posts on this page:** 6\
**Page:** 1

<div class="post-metadata">

**Author:** ![rdmpage](https://yyz1.discourse-cdn.com/flex033/user_avatar/forum.biodiversitylibrary.org/rdmpage/32/8_2.png) [@rdmpage](https://forum.biodiversitylibrary.org/u/rdmpage)\
**Post date:** [March 25, 2026, 4:29pm UTC](https://forum.biodiversitylibrary.org/t/sota-ocr-state-of-the-art-optical-character-recognition/42/1 "2026-03-25T16:29:41Z")

</div>

I’m hoping to win the prize for most acronyms in a title! This post is a placeholder for tracking the current **s** tate **o** f **t** he **a** rt (SOTA) in OCR. If you come across a tool that might be good for BHL, or indeed has been used, please add it here.

[https://www.datalab.to](https://www.datalab.to)

> **[Mistral OCR | Mistral AI](https://mistral.ai/news/mistral-ocr)**
>
> Introducing the world’s best document understanding API.

---

<div class="post-metadata">

**Author:** ![Pigsonthewing](https://yyz1.discourse-cdn.com/flex033/user_avatar/forum.biodiversitylibrary.org/pigsonthewing/32/50_2.png) [@Pigsonthewing](https://forum.biodiversitylibrary.org/u/Pigsonthewing)\
**Post date:** [March 27, 2026, 9:00pm UTC](https://forum.biodiversitylibrary.org/t/sota-ocr-state-of-the-art-optical-character-recognition/42/2 "2026-03-27T21:00:05Z")

</div>

I’ve heard good things about Transkribus, especially for non-western scripts and handwriting

> **[Transkribus - The AI Platform for Historical Documents](https://www.transkribus.org/)**
>
> The leading AI platform that turns handwritten documents into searchable, interactive knowledge. Trusted by archives, built for everyone.

---

<div class="post-metadata">

**Author:** ![rdmpage](https://yyz1.discourse-cdn.com/flex033/user_avatar/forum.biodiversitylibrary.org/rdmpage/32/8_2.png) [@rdmpage](https://forum.biodiversitylibrary.org/u/rdmpage)\
**Post date:** [March 27, 2026, 9:29pm UTC](https://forum.biodiversitylibrary.org/t/sota-ocr-state-of-the-art-optical-character-recognition/42/3 "2026-03-27T21:29:11Z")

</div>

Yes, although I seem to recall reading somewhere that mainstream LLMs are getting so good that the trained models of Transkribus may be obsolete 🤷‍♂️. Perhaps a version of [http://www.incompleteideas.net/IncIdeas/BitterLesson.html](http://www.incompleteideas.net/IncIdeas/BitterLesson.html)

---

<div class="post-metadata">

**Author:** ![rdmpage](https://yyz1.discourse-cdn.com/flex033/user_avatar/forum.biodiversitylibrary.org/rdmpage/32/8_2.png) [@rdmpage](https://forum.biodiversitylibrary.org/u/rdmpage)\
**Post date:** [March 27, 2026, 10:18pm UTC](https://forum.biodiversitylibrary.org/t/sota-ocr-state-of-the-art-optical-character-recognition/42/4 "2026-03-27T22:18:05Z")

</div>

Perhaps what we should do is have a collection of BHL pages that represent all the sorts of pages BHL has (e.g., ranging from handwritten, old fonts, non-Latin scripts, through to modern journals) and use this to benchmark new OCR tools as they come out. Maybe the tech team has such a benchmark already (@cajunjoel?). Given that we are starting to see LLMs embedded in web browsers, maybe OCR text becomes something users do for themselves? The OCR built into my Mac is often better than what BHL provides.

---

<div class="post-metadata">

**Author:** ![Pigsonthewing](https://yyz1.discourse-cdn.com/flex033/user_avatar/forum.biodiversitylibrary.org/pigsonthewing/32/50_2.png) [@Pigsonthewing](https://forum.biodiversitylibrary.org/u/Pigsonthewing)\
**Post date:** [April 2, 2026, 5:19pm UTC](https://forum.biodiversitylibrary.org/t/sota-ocr-state-of-the-art-optical-character-recognition/42/5 "2026-04-02T17:19:52Z")

</div>

> [@rdmpage](#):
>
> maybe OCR text becomes something users do for themselves

That won’t help with searchability.

---

<div class="post-metadata">

**Author:** ![rdmpage](https://yyz1.discourse-cdn.com/flex033/user_avatar/forum.biodiversitylibrary.org/rdmpage/32/8_2.png) [@rdmpage](https://forum.biodiversitylibrary.org/u/rdmpage)\
**Post date:** [April 2, 2026, 7:39pm UTC](https://forum.biodiversitylibrary.org/t/sota-ocr-state-of-the-art-optical-character-recognition/42/6 "2026-04-02T19:39:08Z")

</div>

Yes, not unless we can feed that corrected text back to BHL, or that text is used to create annotations (e.g., this page has the species name xxx).

I guess the interesting question here is to what extent we focus on a better global experience (e.g., OCR all of BHL) versus enabling a better local experience (better OCR for the page you are looking at, perhaps using an OCR engine of your choice).