OCR PDF

Searchable PDF or plain text

Searchable PDF keeps every page as it is and adds the hidden text on top; fonts, images and layout are untouched and the file grows by a few kilobytes per page. Plain text gives you the recognised words in a .txt file with a marker between pages, for pasting into a document or a spreadsheet. Pages that already contain real text are detected and left alone, so running OCR on a mixed document only touches the scanned pages.

• Searchable PDF: same look, text you can select, search and copy
• Plain text: words only, one block per page
• Pages with existing text are skipped, not duplicated

Languages

English, French, German, Spanish, Italian and Portuguese are available. Choosing the right language matters: the engine uses a dictionary and letter shapes specific to it, so accents such as é, ñ and ü and long compound words come out correctly. The data for a language is one to three megabytes, downloaded from PaperKit the first time and cached by your browser afterwards. Text in other alphabets is not supported yet.

What recognises well and what does not

Printed text at ordinary sizes, scanned straight and with decent contrast, is recognised almost perfectly. Results drop with small print under about eight points, heavy skew, coloured backgrounds, stamps over text and low-resolution photos. Handwriting is not recognised. If a phone photo is the source, run it through Scan to PDF first: straightening the page and the Document filter raise accuracy more than anything else.

• Best: 200 to 300 dpi scans of printed pages
• Weaker: phone photos at an angle, faxes, tiny footnotes
• Not supported: handwriting, non-Latin scripts

How the hidden text lines up

For every word the engine reports a box and a baseline. PaperKit writes the word in an invisible render mode at that baseline, scaled to the height of its line and stretched to the width of its box, so a selection drawn over the picture covers the right words. On rotated pages the text follows the page rotation. The original page content is not re-encoded, so image quality is exactly what it was.

Speed and size

Recognition runs on your processor: expect a few seconds per page on a laptop and longer on a phone, with a progress line while it works. A fifty-page scan takes a few minutes. Nothing is sent anywhere, so there is no page limit and no queue. To shrink the result afterwards use Compress PDF; the text layer survives compression.

Private by design

Online OCR services read your document on their servers, which is a poor fit for contracts, medical letters and identity papers. Here the recognition engine is a WebAssembly module in your tab, the language files come from this site, and the PDF is assembled locally with pdf-lib. Closing the tab removes everything.