OCR a PDF withoutuploading it anywhere

A scanned PDF is a stack of photographs of paper. There are no letters in the file for a computer to find, which is why Ctrl+F comes up empty and copying gives you nothing.

OCR fixes that. It looks at the picture of each page, works out which characters are in it, and records where each word sits. Those words are then laid over the page invisibly, so the document looks identical and behaves like a normal text file.

Drop the file in. The tool first checks whether the PDF already carries real text, and tells you if it does, so you do not spend ten minutes recognizing something that was already searchable.

Click Make searchable and watch the page counter. When it finishes, click Download and you get a copy with -searchable in the name. The Cancel link stops the run at any point.

Nearly every OCR site works by taking your file, reading it on their servers, and holding a copy for a while afterwards. For a passport scan, a tax notice, or a medical letter, that is a strange thing to agree to.

Here the recognition engine is downloaded into your browser and does the reading on your own machine. The first run has an extra wait while that engine loads, and after that it is cached. Nothing about the document goes anywhere.

English, and only English. There is no language menu, because one language pack is what the tool loads.

Other languages written in the Latin alphabet often come through partly readable, since the letter shapes match, though accented words and unusual spellings will be wrong more often. Documents in Devanagari, Arabic, Chinese, Japanese, or Cyrillic will not produce useful text, and characters outside the Latin set are left out of the text layer entirely.

Aim for a flat, straight scan around 300 dots per inch, with dark text on a light background. That is what a normal office scanner produces on its default setting.

The things that hurt most: a page photographed at an angle, a shadow across the paper, faint or toner-starved print, a fax that has been copied a few times, and very small type. If a page comes back as nonsense, the fix is almost always a better scan rather than another run.

Not reliably. The recognizer is built for printed type, and handwriting comes back as scattered letters at best. Neat block capitals on a form sometimes work. Cursive notes in a margin will not, and a signature is a picture as far as OCR is concerned.

Each page is rendered, cleaned up, and read on its own, and recognition is the slow part. Plan for a few seconds per page on a laptop, so a fifty page scan is a coffee break rather than an instant result.

There is no page limit, which is the upside of the work happening on your machine. Free does cap finished files at 10 a day across every tool. If a document is very long, split it and run the parts you actually need.

That means it found real text in the file, so search and copy should already work. You can still run OCR if the existing text layer is wrong, garbled, or covers only part of the document. Just know that the rerun replaces the pages with pictures of themselves, so only do it when the text you have is worse than the text you would get.

Every page in the download is an image drawn at about 144 dots per inch, with the recognized words sitting invisibly on top. For a scan, that is a fair swap and the page looks the way it always did.

Two things follow from it. The file is often larger than the one you started with, and any real text or crisp vector graphics in the original are flattened into the picture. That is the reason to point this tool at scans and leave normal text PDFs alone.

OCR is very good at ordinary prose and much weaker on dense tables, columns of figures, and faint print. A 1 read as a 7, or a 0 as an O, changes an amount. Skim the recognized text against the page for anything that ends up in an accounting system or a legal filing.

Once the layer is in place you can select a paragraph and paste it. If you want the whole document as plain words instead of selecting page by page, run the searchable copy through the text extractor and download a .txt file.

This is the everyday reason to run OCR: an old contract, a stack of receipts, an archived case file where the thing you need is on one page out of two hundred. After the run, Ctrl+F finds it, and a screen reader can read the page aloud.

Recognized text is selectable, not editable, because what you see is still a picture of the page. To rework the wording, copy the text out or convert the file, then edit it in a word processor.

A locked PDF cannot be opened for reading, so the upload card will turn it away. Take the password off first, run the recognition on the plain copy, and lock it again afterwards if it needs to travel protected.

It works, and for a couple of pages it is fine. Recognition is demanding work though, so a long document on a phone means a hot device, a flat battery, and a much longer wait than the same job on a laptop.