Free Tool

PDF to Text with OCR

Read the words out of a scanned document. Pages that already have text never leave your browser.

Pages that already contain text are read in your browser and never uploaded.

Why this tool works the way it does

Optical character recognition is the one job in this site's PDF collection that a browser cannot do. Everything else here — merging, splitting, rotating, numbering, watermarking, converting to images — runs entirely on your own machine. Recognition needs a trained model and real processing, so it runs on a server, and that means pages have to be sent.

So the tool is built to send as little as it possibly can. When you choose a file, every page is examined in your browser for a text layer. Pages that have one are read locally, immediately, at no cost and with nothing transmitted. Only the pages that turn out to be pictures are gathered into a new PDF and sent for recognition. The panel shows those numbers before anything happens, so you can see exactly what is about to leave the device.

The common cases this produces are worth spelling out:

  • An ordinary digital PDF — nothing is sent at all. Every page is read in the browser and the tool says so.
  • A report with two scanned inserts in forty pages — two pages are sent, not forty.
  • A fully scanned document — all of it is sent, because there is no other way to read it.

Which tool should you use?

If your PDF is a normal digital document, PDF to Text does the whole job in your browser with nothing sent anywhere, and is the better choice. Come here when that tool returns nothing or when your file is a scan.

Getting good results

Recognition quality depends far more on the scan than on the software. On a clean scan of printed text it typically reads well above 95 per cent of characters correctly. What pushes it down:

ProblemEffectWhat helps
Low resolutionLetters blur into each otherRescan at 300 DPI; below 200 accuracy drops sharply
Skewed or curved pagesLines are missed entirelyStraighten the page, or photograph it flat
Photograph at an angleDistorted charactersShoot square to the page with even light
Faint or dot-matrix printCharacters guessedIncrease contrast before scanning
Wrong language selectedPlausible but wrong wordsName the language; use eng+hin for mixed text
HandwritingLargely unreadableNo reliable fix — this is printed-text recognition

The language setting matters more than people expect. Recognition uses it to choose between characters that look nearly identical, so a document read as the wrong language returns real words that are not the ones on the page — a failure that is much harder to spot than obvious nonsense.

Always read the result

Recognition is a very good guess, not a transcription. A digit misread in an invoice or a name misread in a record is the kind of error that survives right through to whatever you paste it into. The text comes back editable here for that reason: read it against the original before you rely on it.

Plain text or Markdown

Plain text is the straightforward option: the words, in reading order, ready to search or paste. Markdown additionally keeps the headings and table structure that recognition could detect, which is worth having when the scan is a structured document rather than continuous prose.

Questions

Is my document uploaded?

Only the pages that need recognition, and often none at all. Every page is checked in your browser first: any page carrying a real text layer is read locally and never leaves your device. Pages that turn out to be images are collected into a separate file and sent for recognition, because reading letters out of a picture is not something a browser can do. The page tells you exactly how many pages will be sent before anything happens.

What is the difference between this and the ordinary PDF to Text tool?

The ordinary tool reads the text layer a PDF already contains, entirely in your browser, and returns nothing for a scan. This one adds recognition for pages that have no text layer. If your document is a normal digital PDF, the other tool does the whole job locally and is the one to use.

How accurate is the recognition?

On a clean scan of printed text at a reasonable resolution, very good — typically well above 95 per cent of characters. Accuracy falls with low resolution, skewed pages, photographs taken at an angle, faint or dot-matrix printing, unusual typefaces and handwriting, which is not reliably readable at all. Always read the result before trusting it.

Which languages are supported?

Around thirty, including English, the major European languages, Hindi, Arabic, Chinese, Japanese and Korean. Naming the right one materially improves accuracy, because recognition uses the language to resolve characters that look alike. A document mixing two scripts can name both, for example eng+hin.

How long does it take?

Recognition runs as a background job, usually a few seconds per page. A long or dense document can take a minute or more. The page polls for the result and shows progress rather than leaving you guessing.

Will the layout be preserved?

Reading order is preserved and the Markdown option keeps headings and table structure where it can detect them. A precise visual reproduction is not the goal: what comes back is the text of the document, ready to search, edit or paste elsewhere.