Translate a PDF to Word — layout kept, checked side by side by you
A translated Word file that keeps each page’s layout, a side-by-side review of every passage, and a translator’s statement for the person who checked it to sign.
Free: a translation counts as 1 of 3 daily jobs, scanned pages use the 5 free OCR pages a day, and files go up to 25 MB. The Word files are unbranded; the statement PDF carries a visible Made with Bindery line. Pro removes the caps and the line.
Prefer the full studio? Open Translate PDF in the app →
Bindery reads the PDF in your browser — its text layer, or the words OCR finds on a scanned page — and rebuilds each page as paragraphs, headings and tables, the way PDF → Word does. Each passage is translated and put back where it came from, so the Word file has one section per source page with the page size, margins and columns of the original, and longer text runs onto an extra page rather than being cut. Page numbers and passages without letters, such as amounts and reference numbers, are left exactly as they are; headers and footers that repeat on every page, and any passage that appears twice, are translated once.
Where the text goes depends on the engine, and the screen says which one is running before you start. In Chrome on a computer, Chrome’s built-in translator works on your device; the first use of a language pair may download a language pack. Edge’s translator is shown as your browser’s, not as on-device, until Bindery has checked it. On other browsers, or by choice, you can use your own AI key — the settings Ask this PDF uses — and then the full text of the pages you translate goes to the provider you chose, only after you tick consent for that job, and never to Bindery. The PDF file itself is not uploaded either way. What comes out is a machine draft for a person to check: Bindery does not certify, notarize or swear translations.
How it works
- Drop the PDF and open Translate PDF — the source language is detected and can be changed; choose the language to translate into
- Press Translate: Chrome translates on your device, or you tick consent and your own key sends the text to your provider; then check each page side by side and fix any flagged passage
- Download the translated .docx, the side-by-side review, or the translation packet with a statement for the translator to sign
Which engine translates, and where the text goes
| Engine | Where it works | What leaves your device |
|---|---|---|
| Chrome’s built-in translator | Chrome 138 or later on a computer; not Chrome on a phone | None of the text. Chrome translates on your device; the first use of a language pair may download a language pack, which Chrome keeps and shares with other sites. |
| Microsoft Edge’s built-in translator | Edge 148 or later on a computer | Not yet confirmed. Microsoft describes the models as on-device, but until Bindery has checked that Edge never hands text to a cloud service, the screen labels it “Translated by your browser (Microsoft Edge)” and Bindery does not call it on-device. |
| Your own AI key | Any browser, phones included | The full text of the pages you translate, with any text read from scanned pages, to the provider you configured — OpenAI, Groq, OpenRouter or Mistral — after you tick consent for that job. That provider’s terms apply; your key stays in this browser. |
| Ollama on your own computer | Any browser on that computer | Nothing beyond the computer: the text goes to the model server on it, at localhost:11434. |
Bindery sends only to api.openai.com, api.groq.com, openrouter.ai, api.mistral.ai, and localhost:11434 or 127.0.0.1:11434 for Ollama: a base URL anywhere else, Bindery’s own domain included, is refused before any request is made. The own-key settings are the ones Ask this PDF uses, so changing them for one changes them for both. Firefox and Safari have no built-in translator, and Chrome on a phone has none either; there, your own key or a local Ollama is the way to translate. Which pairs Chrome can do is asked of the browser before you start, not assumed from a list.
What carries over into the Word file
- Layout: one Word section per source page, with its page size, margins and one or two columns; longer text runs onto an extra Word page rather than being cut.
- Structure: headings, paragraphs and tables come back as Word headings, paragraphs and tables, and each table cell is translated on its own.
- Labels and values: on forms and official records, a line such as “Date of birth: 4 March 1990” is split at the colon, so the label and the value are translated separately.
- Numbers and codes: passages with no letters — amounts, dates written in digits, reference numbers — and page numbers are left exactly as they are, and so is a URL or an email address on its own.
- Styles: bold or italic inside a paragraph collapses to the paragraph’s main style, except in a label and its value.
- Images: pictures stay where they were, untranslated, and the review says how many there are, because seals, stamps and signatures may need a bracketed description. On a page read with OCR, the picture of the scan itself is left out, since its text is now there in the target language.
The review, passage by passage
Before anything is written, every page is shown with the source on the left and the translation on the right, in a box you can edit. Each passage is checked, and anything doubtful is flagged: a number in the source that is missing or different in the translation (1.234,56 and 1,234.56 count as the same number), a translation identical to the source, an empty one, a length out of proportion for the two languages, a passage that came back in the wrong language, and words OCR read with low confidence. Only flagged hides the rest.
A flag never blocks the download; a passage that failed to translate does, until you translate it again or keep the original text. Translate again sends a passage back: to Chrome’s translator only when it failed, or to your own key, after you tick consent again if the job used Chrome. Your edits go into every file you download, and the same two columns download as a Word file — review-side-by-side-report.docx for report.pdf.
The translation packet and the statement
For a translation that has to be handed in, the packet puts everything in one ZIP: the original PDF exactly as you dropped it, the translated .docx, a translator’s statement as .docx — and as PDF when its text can be drawn with the standard fonts — the side-by-side review if you want it, and a README.txt listing what is inside, the original’s SHA-256 fingerprint, and every engine and host that produced text.
The statement is in English. It is a declaration by the person who checked the translation: that they are competent in both languages, that the attached translation of the named document was made or fully reviewed by them, and that it is complete and accurate to the best of their knowledge; when a machine draft was used, it says so and that they reviewed all of it. They fill in their name, signature, date and contact. The name is never filled in from an account, and a name containing “Bindery” is refused. On the done screen, Sign the statement opens the statement PDF in Sign.
Bindery does not certify, notarize or swear translations, and the footer of the statement says so. Recipients set their own rules: some ask for a certified or sworn translator, notarization, or the documents in a different order, and the README says that Bindery provides none of these.
Honest limits
- A machine draft. Machine translation misses context, idiom and terms of art, and can drop a word or a number. Read every passage against the original before relying on it; the flags find the doubtful ones, not all of them.
- No translated PDF yet. A PDF written from the translation would lose the page-by-page match with the original that the Word file keeps, and the standard PDF fonts Bindery can use without downloading a font (Helvetica) draw only a basic Western European character set. Open the Word file in Word, Pages or LibreOffice and save it as PDF there.
- Text in pictures stays as it is. Words inside images, form-field values, comments and annotations are not translated.
- Scans need an OCR language. A scanned page is read in one of the 11 OCR languages; a scanned page in any other language is skipped and named in the review, not misread.
- Right-to-left sources. With an Arabic or Hebrew PDF, check the word order of the first page in the review before relying on the rest.
- Keep the tab open. Translation runs in this tab, and nothing is kept once it is closed. While it is open, finished passages are kept in memory, so running the same file and pair again does not redo them.
- It needs the network. Unlike the other tools, Translate PDF needs a connection: your own key needs it every time, and Chrome needs it once per language pair, to download its language pack.
Questions people ask
Is my PDF uploaded to translate it?
The file is not: it is opened, read and written in your browser tab. What happens to its text depends on the engine. Chrome’s built-in translator on a computer works on your device. Your own AI key sends the full text of the pages you translate to the provider you configured — OpenAI, Groq, OpenRouter, Mistral, or an Ollama server on your own computer — and only after you tick consent for that job. Your key stays in this browser, and no text is sent to Bindery.
Who vouches for the translation?
The person who checks it, not Bindery. Bindery does not certify, notarize or swear translations, and it has not checked yours: it gives you a machine draft and a side-by-side review to correct it. The translation packet adds a translator’s statement for that person to fill in and sign in their own name. If a court, a university or an immigration office asks for a certified or sworn translator, or for notarization, the statement does not replace that.
Which languages work?
Translate PDF offers 39 languages, from Arabic, Bengali and Bulgarian to Turkish, Ukrainian and Vietnamese. In Chrome, the pair you choose is checked with the browser before you start, and a pair Chrome cannot do can go through your own key instead. The Word file works for all of them: Arabic and Hebrew come out right to left, and for every non-Latin script the original fonts are dropped so Word picks one that can show it.
Does it work on scanned PDFs?
Yes, in the 11 languages Bindery’s OCR reads: English, Spanish, French, German, Portuguese, Italian, Arabic, Simplified Chinese, Japanese, Russian and Hindi. Scanned pages are read on your device and use the free allowance of 5 OCR pages a day; Pro has no limit. On a page read with OCR, the picture of the scan is left out of the Word file, because its text is now there in the target language. A scanned page in any other language is skipped and named in the review, not misread.
What does the free plan include?
All of it. A translation counts as 1 of your 3 free jobs a day, however many pages it has, and building the Word files, the review and the packet afterwards does not count again. Files can be up to 25 MB, and scanned pages use the 5 free OCR pages a day. The Word files are never branded; on the free plan the statement PDF carries the small “Made with Bindery” line, kept clear of the signature, and the original PDF in the packet is left untouched. Pro removes the caps and the line.
How do I use my own AI key, or Ollama?
The own-key settings are the ones Ask this PDF uses: an OpenAI-compatible base URL, a key and a model, stored in this browser. OpenAI, Groq, OpenRouter and Mistral work at their own addresses. For Ollama, start it with OLLAMA_ORIGINS=https://binderypdf.com so the browser may reach it, and allow Chrome to connect to apps on this device if it asks; passages go to it in smaller batches, because Ollama’s default context is small. Every job asks you to tick consent, naming the provider and how much text it will receive, before anything is sent.
What do the flags in the review mean?
Each one points at something to look at: a number in the source that is missing or different in the translation, a passage left identical to the source or empty, a length out of proportion for the two languages, a passage that came back in the wrong language — often a model’s refusal or an untranslated paragraph — or words OCR read with low confidence. Flags never block the download. A passage that failed to translate does, until you translate it again or keep the original text.