OCR a scanned PDF — without uploading

OCR is the one category where local processing is genuinely rare. Character recognition is computationally heavy, so nearly every OCR service runs it on their servers — which means uploading the document. And the documents people need OCR for are almost always scans of paperwork: contracts, bank statements, medical letters, ID documents, historical records.

This runs Tesseract compiled to WebAssembly directly in your browser. The engine and its English language data are served from this site and cached on first use, and after that recognition happens on your own CPU. Your scan is never transmitted, because there is no backend here to transmit it to. Once cached, OCR works with the network disconnected.

A PDF can hold text in two completely different ways. A document exported from Word or a web page contains real characters, and those can simply be read out — that is what Extract text from a PDF does, instantly and perfectly.

A scanned document contains no characters at all. It is a photograph of a page, and the text you see is just an arrangement of dark pixels. OCR is the process of examining those pixels and deciding which letters they represent. It is slower and imperfect by nature, which is why you should always try plain extraction first and only reach for OCR when it comes back empty.

The quick test: if you can select individual words in a PDF viewer, use text extraction. If your cursor only draws a rectangle over the page, you need OCR.

Recognition quality depends almost entirely on the input image. Scans at 300 dpi, straight rather than skewed, with good contrast and no shadow across the page, produce dramatically better results than a hurried phone photograph taken at an angle in poor light.

• Scan at 300 dpi where you can — higher is rarely better and is much slower.
• Keep the page square to the camera or scanner bed; skew hurts more than resolution.
• Plain printed text is recognised well. Handwriting mostly is not.
• Unusual fonts, heavy stylisation, and low-contrast scans all reduce accuracy.

Expect a few seconds per page, depending on your machine — real character recognition is heavy work, and because it runs locally a faster device means faster OCR. English is supported today; each additional language needs its own recognition data file, so more may follow. Always proofread the output against the original before relying on it, particularly for numbers, where a mistaken digit is easy to miss and expensive to inherit.