Boox handwriting recognition
Why is Boox handwriting recognition inaccurate, and what are the alternatives?
Updated
Short answer: built-in recognition reads your strokes letter by letter, in one language, with little sense of the page around them. Neat single-language prose comes out well; quick notes, abbreviations, names and mixed languages don't.
A few settings help. The real alternative is a vision model that reads the whole page the way a person does, which is what Margen does for every page you write.
Why does Boox recognition get words wrong?
- The language setting. Recognition uses a model and a dictionary per language. Write in a different language than the one set, or mix two on a page, and it forces your words into the wrong vocabulary.
- Words outside the dictionary. Names, product codes, jargon and abbreviations ("w/", "mtg", "→ Ana") get "corrected" into ordinary words.
- Strokes, not the page. It reads ink in sequence. Arrows, words in the margin, columns, crossings-out and a list drawn next to a sketch confuse the reading order.
- Speed. Fast writing joins letters and drops strokes. You can still read it; a stroke-based recognizer often can't.
How can I make Boox recognition more accurate?
- Set the recognition language to the one you write in, per note if you switch.
- Convert smaller selections: a paragraph at a time, not a whole page with sketches.
- Write on lines or a template with ruled lines; keep margin notes for things you don't need converted.
- Update the firmware: Boox has kept improving its recognition.
If your handwriting is neat and in one language, that may be all you need. It's free and works offline.
What are the alternatives?
| Alternative | How it reads | Trade-off |
|---|---|---|
| Upload pages to Claude or ChatGPT | A vision model reads the page image | Good on messy writing; you export and upload each note by hand |
| Your own pipeline (export PDFs, run them through a model) | A vision model, automated by you | Flexible; takes setup and an API key, and you maintain it |
| Margen | A vision model reads each page you write, automatically | Nothing to set up; notes written in Margen, not Boox Notes; needs a cloud copy |
How Margen reads handwriting
Margen is a notes app for Boox tablets. When you close a page, an image of it goes to a vision model, which returns the text in reading order, a short title and a one-line summary. Because it sees the whole page, it handles mixed languages, abbreviations, lists, arrows and margin notes far more like a person would. It still misreads a word now and then, so the page image stays next to the text, and Claude can check it.
"It read my handwriting very accurately and also pointed out the one word that it thought may have been incorrectly interpreted."
Margen early-access user
The text is read from Claude (any plan) or ChatGPT (paid), connected once. Margen is free for up to 100 new notes a month; Pro, $49 a year, has no limit. For a side-by-side with Boox's own app, see Boox Notes vs Margen.
Questions
Why is Boox handwriting recognition inaccurate?
It recognizes strokes letter by letter in one language, with a dictionary. Mixed languages, names, abbreviations, fast writing and page layout (arrows, margins, columns) are where it makes mistakes.
How can I improve Boox handwriting recognition?
Set the recognition language to the one you write in, convert smaller selections, write on ruled lines, and keep the firmware up to date.
What is more accurate than Boox handwriting recognition?
Vision models, which read the page image like a person: Claude or ChatGPT on an uploaded page, or Margen, which sends every page you write on a Boox to a vision model automatically.
Does Margen work offline like Boox recognition?
Writing works offline, but transcription needs a connection: each page is read by a vision model in the cloud.