Pull the text out of a screenshot, a photo of a receipt, a slide someone shared as a picture, or anything else where the words are trapped in an image. Recognition runs in your browser through Tesseract compiled to WebAssembly, so the picture never leaves your device — which matters when the thing you are reading is a payslip, a contract or a medical letter. Paste a screenshot straight in with Ctrl-V, or drop a whole folder. Fifteen languages are available, and picking only the ones actually present matters: every extra language gives the recogniser more characters to confuse, so a scattergun selection usually makes the result worse. The text comes back in an editable box with wrapped lines rejoined into real paragraphs, and with a confidence figure — because a low-confidence result that still reads like prose is more dangerous than an empty one, since nothing in it warns you that a digit or a name has been invented.
| Source | Expected result |
|---|---|
| Screenshots of text | Excellent — clean, high-contrast pixels |
| Scanned printed pages | Very good if straight and in focus |
| Photographed documents | Good when square, sharp and evenly lit |
| Receipts | Usually good; check digits carefully |
| Text over busy images | Mixed — low contrast is the enemy |
| Handwriting | Poor — Tesseract is trained on printed type |
No. Recognition runs in your browser. The language model is downloaded once from a public CDN and cached, but your image is never transmitted — which is the point when the thing you are reading is a payslip, a contract or a medical letter.
The language model is 10 to 15 megabytes and has to be fetched the first time you use it. Your browser caches it, so every run after that starts immediately, and further images with the same language reuse the already-loaded worker.
Straighten the page and fill the frame. OCR is far more sensitive to focus, skew and contrast than to resolution — a sharp, square, well-lit photo at a modest size beats a large blurry one every time. Reaching for more megapixels rarely helps.
Only if several are genuinely present. Each extra language adds characters the recogniser can confuse, so selecting more than you need usually lowers accuracy. One language is the right answer for most documents.
It is Tesseract's own average certainty across the recognised words. Below 80 percent, proofread numbers and names carefully; below 60, treat the whole result as a draft. OCR fails by inventing plausible text rather than by leaving gaps, so the figure is worth reading.
Not reliably. Tesseract is trained on printed type, and handwriting results range from poor to unusable. Printed text, including photographs of it, is what it does well.
Yes — the output is an editable text box, not a read-only panel. Correct it in place, then copy or download.
Tesseract compiled to WebAssembly does the recognition, with the language model fetched once from a public CDN and cached by the browser thereafter — the image itself never leaves the page. The worker is created once per language selection and reused across every file, because initialising a worker re-downloads and re-parses 10 to 15 megabytes of training data, which costs far more than recognising another image. Language choice matters more than most people expect: each extra language gives the recogniser more characters to confuse, so selecting several usually lowers accuracy rather than raising it, and the interface says so rather than encouraging a scattergun selection. Raw Tesseract output breaks a line wherever the source image did, which means a paragraph arrives as a stack of short fragments that no longer wrap when pasted anywhere else. The cleanup pass rejoins those, treating a blank line or a line ending in sentence punctuation as a genuine break, and rejoins words split by a trailing hyphen without leaving the hyphen behind. Form feeds, which Tesseract emits between pages, become paragraph breaks. Mean per-word confidence is reported alongside the text, and a result below eighty percent is flagged — a low-confidence result that still looks like prose is considerably more dangerous than an empty one, because nothing about it signals that the numbers and names in it may be wrong. Output lands in an editable box, since OCR nearly always needs a correction or two.