Read text from a scan

A PDF made by a scanner or a phone camera contains a picture of a page, not the words on it. Selecting text does nothing, search finds nothing, and a screen reader announces nothing, because as far as the file is concerned there is only an image. Optical character recognition looks at the shapes and works out which letters they are.

Our PDF to text tool pulls out text that is genuinely embedded in a document and tells you when a PDF is a scan. This page is what you use when it is.

Almost every OCR service works by uploading your document to a server. That is a meaningful thing to be asked to do with the sort of paper people usually scan: contracts, medical letters, bank statements, passports, court filings, tax records.

Here the recognition engine itself is downloaded to your browser — about six megabytes the first time, then cached — and the document is read where it already is. The engine is Tesseract, the same open-source engine behind a great many paid OCR products, compiled to WebAssembly. Nothing is sent anywhere at any point, which you can confirm in your browser's Network tab.

Reliable: printed text in English, reasonably straight, at a sensible resolution. A flatbed scan at 200 or 300 DPI is close to ideal. A phone photo of a printed page, held square and in decent light, is usually fine.

Unreliable: handwriting — this engine is trained on print and will produce confident nonsense from cursive. Also heavily stylised type, very low-resolution images, photographs taken at an angle, tables where the columns matter, and text over busy backgrounds. Multi-column layouts are read in the order the engine guesses, which is often but not always the order you would read them.

Not supported yet: languages other than English. The language data for each one is a separate few megabytes, and shipping every language to everybody is not a reasonable trade. If you need another, say so and it can be added.

OCR output always contains mistakes. Ones and lowercase Ls, zeros and capital Os, and rn against m are the classic confusions, and they survive a casual read. The confidence figure shown with the result is the engine's own average across the page and is a rough guide, not a guarantee — a page can score well overall and still have a wrong digit in the one number you care about.

If you are transcribing an account number, a dosage, a date or a legal citation, read it against the original. That advice applies to every OCR product, including the expensive ones; the difference is that this page says so.