How to Translate a Scanned Document (PDF or Image), Step by Step

A scanned document is just a photo of a page — a contract you scanned, a certificate someone emailed you, a phone snapshot of a form. You can read it, but you can't select the text, and that's exactly why translating it is more awkward than it should be. Search for how to do it and you'll mostly find two kinds of advice: "extract the text and paste it into ChatGPT," or a forum thread that says "open the PDF, open Word, and retype it." Both work in a narrow sense — and both quietly destroy your formatting.

This guide lays out the three real ways to translate a scanned document (whether it's a PDF or an image), what each one costs you in time and layout, and how to pick. "Document" and "PDF" mean the same thing here — most scanned documents are saved as PDFs, and the steps are identical either way.

First, why you can't just translate it directly

A normal document created on a computer has a text layer — real, selectable characters underneath what you see. A scanned document has none. It's an image, so to any translator every word is just pixels. Before anything can translate it, software has to look at the picture and work out what the characters are. That step is OCR (optical character recognition), and it's the part most quick methods either skip or do badly.

Method 1 — The ChatGPT / copy-paste route

This is the most-shared advice online: run the scan through an OCR tool to pull the text out, copy that text, paste it into ChatGPT (or DeepL, or Google Translate), and ask for a translation.

1. 1OCR the scan. Use Acrobat's "Recognize Text", a free OCR site, or ChatGPT's file upload to turn the image into selectable text.
2. 2Copy the recognized text. Paste it into your translator of choice and ask for the target language.
3. 3Get plain translated text back. You'll have the meaning — as a flat block of text, with none of the original structure.

The catch is in step three. The moment the text leaves the page, the layout is gone: tables collapse into runs of words, columns merge, and stamps, signatures, and field labels lose their position. For a single paragraph where you only need the gist, that's fine. For a contract, certificate, or transcript — anything where which value sits in which box matters — you'll spend 30–60 minutes rebuilding the formatting by hand. (It's also why these how-to pages tend to have very short visit times: people try it, hit the reformatting wall, and leave.)

Method 2 — OCR into a Word file, then translate

A small step up: instead of copying raw text, export the OCR result to a Word document, then use Word's or Google Docs' built-in "Translate document" feature. This keeps simple paragraph formatting and is reasonable for a mostly-text page. But OCR-to-Word reconstruction is rough on anything structured — multi-column layouts, tables, and official forms still come out misaligned, and you're back to manual cleanup. It also depends entirely on the OCR being accurate; a low-resolution or skewed scan produces fluent-looking nonsense.

Method 3 — A one-step scanned-document translator

The reason the first two methods are painful is that they treat OCR, translation, and re-formatting as three separate chores you stitch together. A purpose-built tool does all three in one pass. Reglyph reads the scanned document with OCR, erases the original text from the page, translates it, and re-typesets the translation back into the exact same positions — so tables, columns, figures, stamps, and numbers stay where they were. You upload the scan as-is and download a finished, layout-correct document. No copy-paste, no Word round-trip, no rebuilding the formatting.

1. 1Upload the scan as-is. A scanned PDF, or even a phone photo of the page. No text layer needed and no separate OCR step — the thing that stops ChatGPT and Google Translate is handled for you.
2. 2It OCRs, translates, and re-typesets. The source text is recognized, translated to your target language, and laid back onto the page in place of the original.
3. 3Download with the layout intact. Real typeset text in the original positions — not a flat block, not an image with words pasted on top. The first 3 pages are free.

Which method should you use?

• One short, plain-text page and you only need the meaning → the ChatGPT / copy-paste route is fine.
• A mostly-text page where rough formatting is acceptable → OCR into Word and translate there.
• A contract, certificate, transcript, or anything with tables, columns, or stamps → a one-step tool that preserves the layout will save you the most time and avoid manual rebuilding.

The short version

You can't translate a scanned document directly because it's an image with no text layer — it needs OCR first. The copy-paste-into-ChatGPT method everyone recommends gets you the words but throws away the layout, so you rebuild it by hand. If your document is structured or official, a tool that does OCR, translation, and re-typesetting in one step gives you back a clean, layout-correct document without the busywork.