Text recognition for pictures of pages: photos, screenshots and scanned PDFs that hold an image rather than real text. It runs in this tab, which matters when the document is a contract or an ID.
A photo of a page, a screenshot, or a scanned PDF. Nothing is sent when you add it.
The recogniser downloads once, about 4 MB, then works with no connection at all. A page takes a few seconds.
Read the result against the original, especially names and numbers, then copy it or save it as a text file.
Printed text on a flat, well lit page is read very accurately. The confidence figure shown with your result is the recogniser's own estimate, and it is worth trusting: above about 88 per cent you are proofreading, and below 70 you are better off retaking the photograph than correcting the output.
Handwriting is a different problem and this does not attempt it. Neither do complicated layouts: multiple columns, tables and forms come out in an order that made sense to the machine and will not to you.
The quality of the picture decides everything. Shadow across the page, a camera held at an angle, and small or faint print are the three things that ruin a result. Two minutes spent retaking the photo in daylight beats twenty minutes of correcting text.
Every other site that does this uploads your document to a server to read it. That is the part worth avoiding when the thing you are scanning is a passport, a payslip or a signed agreement, and it is why this one runs on your own machine instead.
No. The work happens inside this tab, on your own machine. Open your browser's network tab while you use the tool, there is no request carrying your file.
Practical limits depend on the tool, browser and available device memory. Large files can fail even when no explicit limit is shown. Keep an original and use a smaller job if this device cannot finish it.
No, and there is no daily limit either. There is nothing to sign up for.
No. This reads printed text. Handwriting needs a different kind of model and we would rather say so than waste your time.
English at the moment. More can be added, each as its own download.
The recogniser and its English data are about 4 MB and download once. After that they are cached and it works offline.
This uses optical character recognition that runs in your browser. It reads printed English text from a photo, a screenshot or a scanned PDF, and PDF pages are drawn at about 150 dots per inch before they are read. Handwriting is not supported, and only English is installed.
The recognition engine is a download of about 4 MB from this site, fetched the first time and kept on your device after that, so later use works offline and nothing is sent anywhere.
Sharp, straight, evenly lit pictures read best. A tilted or shadowed photo of a page will produce errors, so retake it flat if you can. The tool shows an average confidence: above about 88 percent is dependable, 70 to 87 is reasonable but worth proofreading, and below 70 means the picture needs a better source.
Layout and tables are not preserved, only the words in order, so check names and numbers against the original. If a PDF already contains real text, Extract text gives it back exactly with no errors.