Out of a document

OCR a PDF: read a scanned document

The one every converter sends you away for.

On this device

Choose a scanned PDF

The reader is downloaded the first time you use it, and a page takes under a second on a graphics card, a couple of seconds on one processor thread in Chrome, up to eleven in Safari.

    Opened on this device · never transmitted

    What this tool does, exactly

    Until now every converter here met a scan and said the same true, unhelpful thing: there are no characters in this file. This is the tool that does something about it — each page drawn, then read, on this device — and it gives you three things back: the text, the text as Markdown, and the scan itself with an invisible text layer, so it can be searched.

    What it does
    Draws each page as an image at 200 dpi, finds every line of text on it with a neural detector, reads each line with a neural recogniser (PaddleOCR's PP-OCRv6, run with ONNX Runtime in this tab), and joins the pages into one document. Then it writes the recognised lines invisibly over the original pages, at the positions they were found, so the scan becomes searchable. The scan never leaves the device.
    When it refuses
    When the PDF already has a text layer. Recognising a page whose characters are already there replaces perfect text with a guess — slower, worse, and invisibly so. The tool checks first, says how many characters it found, and sends you to PDF to Markdown or PDF to Word, which read what is really in the file.
    Why 200 dpi
    Pages are drawn at 200 dpi, fixed: it is the resolution the reader was measured at, it keeps ten-point body text well above the size the detector loses, and 300 dpi would cost a long document its memory for very little on printed text.
    What it gives back
    The text in reading order, a Markdown version with rows of cells as table rows, and the searchable PDF — the same pages, with the words laid over them in a font that draws nothing. The searchable PDF is read back through the same PDF library before it is offered, and a page whose layer did not read back is named.
    Which languages
    English and forty-six other Latin-script languages, Chinese and Japanese with one model; Russian, Ukrainian and Belarusian with a second recogniser you choose in the options, an extra eight megabytes.
    What it costs to load
    About seventeen megabytes compressed the first time — the page renderer, the reader and its two models — fetched only when you open this tool and cached by your browser afterwards. It is the heaviest thing on this site, which is exactly why nothing else pays for it — until you open the Pro door under the button, where the Deep reader costs a 660 MB download from Hugging Face, once; the AI OCR page, linked from the door, explains both readers.

    About this tool

    How is this different from PDF to Markdown or PDF to Word?

    Those two read the characters a PDF already contains, which is fast and exact. This one is for the documents that contain none — a scan is a picture of text — and it recognises the shapes instead. If your PDF has a text layer this tool will tell you so and send you to them, because reading what is there always beats guessing at it.

    How accurate is it?

    On a clean, flat scan of printed type, high enough to paste and skim: measured on a synthetic 200 dpi scan, 0.2% of characters wrong in English, 0.7% in Portuguese, 0.6% in Russian. On a phone photograph of a curved page in poor light, low enough that you should read it against the original. The lines the reader was least sure of are listed, so the checking has somewhere to start rather than being a guess about a guess.

    Do I get a searchable PDF back?

    Yes. The recognised lines are written invisibly over the original scan, each at the position it was found, in a font that draws nothing — the same construction Tesseract's PDF output has used for a decade. Any viewer can then search, select and copy the text while showing only the scan. Before the file is offered it is opened again and its text read back, and the result panel says whether every page passed.

    How long does a long document take?

    About 0.6 seconds a page on a graphics card, measured on an Apple M4 Mac; about two seconds a page on one processor thread in Chrome and six to eleven in Safari; plus the seventeen-megabyte download the first time. Pages are drawn, read and released one at a time, so a long document costs time rather than memory. Phones were not measured.

    Which languages does it read?

    English and forty-six other Latin-script languages, Chinese and Japanese with the model that loads by default. Russian, Ukrainian and Belarusian with a second recogniser you pick in the options, because it is an extra eight megabytes that most visitors should not pay for.

    Does the scan leave my device?

    No. The renderer, the reader, the models and the recognition all live in this tab. This matters more here than almost anywhere on the site: the documents people scan are contracts, medical letters, payslips and identity papers, and every free online OCR service receives them. Here they are read and nothing is sent.

    More out of a document tools · All 52 · How they work