Image to Text (OCR)
Drop one or many images (PNG, JPG, WebP, BMP) and pull the text out. Pick a language - or up to three for mixed-language pages - then copy or download all extracted text as a .txt file. Recognition runs entirely in your browser using Tesseract.js with support for 100+ languages.
All OCR runs in your browser via Tesseract.js. Images and extracted text are never uploaded.
How to use this image to text (ocr)
- Drop images onto the box, or click to choose files (PNG, JPG, WebP, BMP). Multiple images at once are fine.
- Pick the document language - or hold to add a second/third language for mixed pages.
- Click Recognize. The first run downloads the language data (~5-10 MB per language, cached for next time).
- Read the extracted text per image with a confidence badge, or copy/download the combined result as a .txt file.
Frequently asked questions
Which languages are supported?
100+ languages via Tesseract.js, including English, German, French, Spanish, Italian, Portuguese, Dutch, Polish, Czech, Russian, Arabic, Hindi, Japanese, Chinese (Simplified and Traditional), Korean, Turkish, Vietnamese, Indonesian, Thai, Ukrainian, Greek, Hebrew, Bengali and Urdu. You can combine up to three languages for mixed-language documents.
How accurate is the OCR?
Tesseract.js typically reaches 85-95% accuracy on clean, high-contrast scans of printed text. Confidence drops on photos with skew, glare, handwriting or low resolution. Each image gets a confidence score so you can spot poor results at a glance.
Why does my handwritten note come out as gibberish?
Tesseract.js is trained on printed text — typeset characters, consistent stroke width, regular spacing. Handwriting violates all three. Cursive is essentially unreadable to it. For handwritten notes use a dedicated handwriting OCR (Apple Notes, Google Lens, OneNote), not a Tesseract-based tool.
Why does the first run take so long but subsequent runs are instant?
The first time you pick a language, your browser downloads the trained model for that language (~5-10 MB) and caches it in IndexedDB. After that the model loads from cache in milliseconds. Mixing languages downloads each one separately.