Make a scanned PDF searchable

It reads the words in the scanned image and stores them in the file as an invisible text layer sitting exactly over the picture of each word. The page looks identical — your scan, unchanged, with its signatures and stamps — but a reader can now search it, select text and copy from it. Nothing visible is replaced or redrawn.

It depends on the scan and on the language, and both are measured rather than guessed at. On clean printed test pages the four Latin languages recover every word we look for, at around 95% confidence. Hindi is noticeably weaker — Devanagari is a harder script, with far more ways to be slightly wrong — and recovers roughly three words in five on the same kind of page. On a faint fax, a skewed photograph or handwriting, all of them degrade sharply. The result panel reports the confidence measured on your own document, which is a better guide than any of these figures. Recognition is a reading of an image, never a transcript, so check anything that matters against the original.

No, and this is the whole reason the tool exists here. Scanned documents are usually the most sensitive things people own — contracts, medical letters, identity papers — and every mainstream OCR service asks you to upload them. Here the recognition engine runs inside your browser and the file never leaves the device.

Because the recognition engine and your chosen language model have to be on your device to run there. Together that is roughly 4.5MB for English and rather less for the other languages; the tool tells you the size before it starts. They are fetched once from this site — never from a third-party CDN — and then stored, so every run afterwards works with no connection at all. If you have already fetched a language, the tool checks and says so rather than warning you about a download you are not about to make.

Yes, once the engine has been stored. Open the tool once while connected and the engine and model are cached; after that you can disconnect entirely and recognition still runs. The very first visit needs the network for that one download.

English, Hindi, Spanish, French and German. Each has its own model, between 0.7MB and 2.8MB, and only the one you pick is downloaded — choosing a language never fetches the other four. The selector can only offer languages whose models are genuinely stored on this site, so it will never present one your browser cannot get. Recognising the wrong language produces confident nonsense, so choosing correctly matters more than it might seem.

Fifty pages at a time. Recognition is far heavier than an ordinary PDF operation — seconds and hundreds of megabytes per page — and an unbounded job on a phone does not fail politely, it takes the tab down. Larger documents can be split first and run in parts.

Yes, and that is the point. Once a document has a text layer, PDF to Word, PDF to Text, PDF to HTML and Highlight all work on it exactly as they would on a document that was born digital. Those tools can also recognise a scan directly if you ask them to.

If nothing at all can be recognised across the pages you selected, you are told so instead of being handed a file that looks processed and is not. A page that is genuinely blank simply contributes nothing, which is correct.