What this does
You’ve got a screenshot of a receipt, a photo of a textbook page, or a scanned PDF page saved as an image, and you need the actual words out of it. Retyping is miserable. This reads the picture and hands you back selectable, editable text.
Under the hood it runs OCR, short for optical character recognition. The engine is Tesseract, an open-source recognizer Google has maintained for years. It scans the pixels, finds shapes that look like letters, and reconstructs them into characters and words. Your image goes up to the server for that pass, then gets wiped automatically after about an hour. Nothing sticks around.
How to use it
- Drop an image on the box or click to browse. Screenshots, scans, and photos all work, as long as the text is legible.
- Pick the language from the dropdown. This matters more than people expect, since the engine uses a per-language model to recognize accented letters and spelling. English is the default.
- Hit Extract text.
- Read the result in the box, fix any stray characters, then Copy or Download .txt.
Heads up on timing. The first time you run a given language, the server loads that language model, so it can take a few seconds. Every run after that is quick.
Where it shines, and where it doesn’t
Clean, high-contrast text is the sweet spot. Black words on a white background, screenshots of articles, slide decks, printed documents, code on a light editor theme. Those come back close to perfect.
Things get shakier fast in a few situations. Handwriting? Don’t count on it. The model is trained on printed type, so cursive and scribbles usually come out as gibberish. Low-resolution images give it almost nothing to work with, so blow the picture up before you shrink it. A photo shot at an angle, with the page tilted or curling, throws off the line detection. And busy backgrounds, watermarks, or text laid over a photo all confuse it.
A quick fix for most of these: crop tight to just the text and make sure it’s well lit and roughly straight. Five extra seconds of cleanup beats fighting a 40 percent confidence score.
About that confidence number. After each run you’ll see a percentage. That’s the engine’s own estimate of how sure it is, averaged across the page. High 90s means trust it. Drop into the 60s and you should proofread every line before you rely on it.
One language at a time, by the way. If your image mixes English and Russian, run it twice, once per language, and stitch the results together. There’s no auto-detect here.
FAQ
Is my image uploaded to a server?
Yep, it has to be. OCR runs server-side, not in your browser, so the file gets sent up for processing. It’s deleted automatically about an hour later, and it isn’t stored or shared.
Which languages are supported?
Nine for now: English, Turkish, Spanish, French, German, Portuguese, Italian, Dutch, and Russian. Choose the one that matches the text in your image for the best accuracy.
Why is the first scan slow and the rest fast?
The server loads the language model into memory on the first request for that language. That warm-up takes a few seconds. Once it’s loaded, later scans skip the wait.
Can it read handwriting?
Realistically, no. The recognizer is built for printed and typed text. Handwriting, especially cursive, comes back unreliable or as nonsense. Stick to typed or printed sources.
The text came out wrong. What can I do?
Usually it’s the image, not the tool. Crop to just the text, use a sharper or higher-resolution version, straighten any tilt, and confirm you picked the right language. Then run it again and check the confidence score.
What image formats can I upload?
Common raster formats like PNG, JPEG, WebP, and similar. If your browser can show it as an image, it’ll upload fine.