Optical character recognition reads text out of pictures of text. A scanned contract, a photographed receipt, a screenshot of an error message — all of them look like documents but contain no text a computer can select, search or copy. OCR examines the shapes of the letters and reconstructs the characters.
This tool runs Tesseract, the long-established open-source OCR engine, compiled to WebAssembly so it executes inside your browser rather than on a server. PDF pages are rendered at twice their native size before recognition, because resolution is the single biggest factor in accuracy. The result is a plain text file, with pages separated by a marker so you can tell where each one begins.
Recognition happens on your own machine. This is unusual for OCR — most services require an upload — and it means payslips, medical letters and identity documents can be read without being transmitted.
.txt file with pages divided by numbered separators.| Factor | Effect on results |
|---|---|
| Source resolution | The dominant factor. A 300 DPI scan reads reliably; a 100 DPI one guesses. |
| Straightness | Skew of more than a couple of degrees costs accuracy quickly. |
| Contrast | Black on white is ideal. Faint photocopies and coloured backgrounds hurt. |
| Typeface | Ordinary serif and sans text is reliable. Handwriting is not supported. |
| Layout | Single-column flows cleanly. Multi-column layouts and tables lose their structure. |
Automatic mode pairs the language of the page you are on with English, so a Korean document is recognised as Korean plus English, a Japanese one as Japanese plus English, and so on. That combination handles the common real-world case of a document in one language containing English product names, addresses or technical terms.
Each language model is a separate download of several megabytes, fetched the first time you use it and cached by the browser afterwards. The first run on a new language therefore takes noticeably longer than the second.
No, and that is the point. Tesseract is compiled to WebAssembly and runs inside your browser tab. Your scans are never uploaded, which makes it safe for documents you would not want to hand to an online service.
On a clean, straight 300 DPI scan of ordinary printed text, accuracy is high — typically well above 95 percent. Faint photocopies, skewed pages, low-resolution photos and unusual typefaces all reduce it noticeably.
No. Tesseract is trained on printed type. Handwritten notes, signatures and cursive annotations are not recognised.
A UTF-8 plain text file. Pages are separated by a numbered marker so you can see where each one starts. Layout, columns, tables and images are not preserved.
The language model has to be downloaded, which is several megabytes. It is cached afterwards, so subsequent documents in the same language start immediately.
No. The output is a separate text file rather than a text layer added back into the PDF. Use it for searching, quoting or feeding into another system.