A scanned PDF contains pictures of pages, not text — which is why copying from it produces nothing. OCR recognises the characters visually. This runs the recognition entirely on your device, so a scanned bank statement or Aadhaar copy is never uploaded anywhere.
Nothing is uploaded. The OCR engine is served from this site and runs in your browser using WebAssembly. Your document is read locally and the recognised text never leaves your device.
Enter your details
Result
Fill in the fields above and your result will appear here.
What the OCR PDF does
The engine is Tesseract, the long-standing open-source OCR project, compiled to WebAssembly. It is about 16 MB including the English language data, downloaded once on first use and then cached by your browser.
That download is the honest cost of doing this without a server. Online OCR services avoid it by uploading your document instead — which for identity documents and financial records is exactly the wrong trade.
How to use this tool
- Select a scanned PDF or a photograph of a document.
- Choose the rendering resolution. 200 DPI suits most scans; use 300 for small print.
- Press Run OCR. The engine downloads on first use, then recognition begins.
- Read the confidence figure, check the preview, and download the text.
- Proofread the result — OCR always needs checking.
How OCR works here
1. PDF pages are rendered to images at the chosen DPI.
(An uploaded image is used directly.)
2. Tesseract, compiled to WebAssembly, processes each image:
binarisation → layout analysis → character recognition
→ dictionary-assisted correction
3. Text and a confidence score are returned per page.
Accuracy depends almost entirely on input quality.
Confidence below about 80% usually means visible errors.
Limited to 30 pages per run, because OCR is slow.
Worked example
Example: a scanned 4-page letter
| Input | 300 DPI scan, clean printed English |
| Confidence | about 92% |
| Time | roughly 10–20 seconds per page |
| Result | accurate text, occasional punctuation errors |
The same document photographed on a phone in poor light might return 60–70% confidence, with numbers and punctuation the first things to go wrong.
What your result means
Confidence above 90% means the text is mostly reliable, with occasional punctuation errors.
70–90% means noticeable errors, particularly in numbers. Proofread carefully.
Below 70% means the input is too poor. Rescan at higher resolution, or clean the image with the scan tool first.
Important considerations
- Input quality dominates everything. A 300 DPI straight scan will always beat a phone photo, whatever settings you use.
- Run poor photographs through the photo to scanned PDF tool first — higher contrast measurably improves recognition.
- English only. Other language data is not bundled, since each language adds another large download.
- Handwriting is not reliably recognised. Tesseract is trained on printed text.
- Multi-column layouts and tables often come out in the wrong reading order.
- Limited to 30 pages per run. Split longer documents first.
- Numbers and punctuation are the most error-prone. Always check figures against the original.
Limitations of this tool
- English only — other languages would each require another large language file.
- Handwritten text is not reliably recognised.
- Complex layouts, tables and multi-column pages often produce jumbled reading order.
- Slower than a server-based service, because everything runs on your device — that is the trade-off for not uploading.
- The engine is about 16 MB on first use, though your browser caches it afterwards.
- Limited to 30 pages per run.
Frequently asked questions
Is my document uploaded for OCR?
No. The OCR engine is served from this site and runs in your browser via WebAssembly. Your document and the recognised text never leave your device — which is the main reason to use this for identity or financial documents.
Why is there a 16 MB download?
That is the OCR engine and the English language data. Services that avoid it do so by uploading your document to their servers instead. It downloads once and your browser caches it.
How accurate is it?
Good for clean printed English — typically above 90% confidence on a proper scan. Noticeably worse on phone photos, unusual fonts and multi-column layouts, and unreliable for handwriting. Always proofread.
Does it work with Hindi?
Not currently. Only English language data is bundled; each additional language would mean another large download.
Why is it slow?
Because recognition runs on your device rather than a server. Ten to twenty seconds per page is normal. That slowness is the direct cost of your document not being uploaded.