ALL IMAGE TOOLS56
Image to Text
Image to Text — Free OCR in Your Browser
Pull the printed words out of a picture without uploading it. Tesseract is compiled to WebAssembly and runs in this tab.
Nothing is enforced, but the ceiling is real: the engine holds the decoded picture and the 4.1 MB English model in wasm memory at the same time, so a very large scan can exhaust the tab before it finishes.
About Image to Text
OCR is the one job on this site that people assume must involve a server, because for twenty years it did. It does not have to. Tesseract — the optical character recognition engine Google has maintained since 2006 — compiles to WebAssembly, and this page runs it in your tab against a picture the browser already has in memory. Nothing is uploaded, which for this particular tool is worth saying plainly: the images people run OCR over are receipts, invoices, passports, prescriptions, contracts and screenshots of private conversations, and every free OCR site that asks you to upload one has taken a copy of it. The cost of doing it this way is a download. The engine and its English language model are about 4.9 MB together, fetched from this site the first time you read an image and not before — the page itself pulls none of it. What comes back is a guess, not a transcript. Tesseract is very good on clean, high-contrast printed text: a screenshot, a PDF page rendered to an image, a flatbed scan of a printed document. It is markedly worse on photographs taken at an angle, on low-resolution images, on text over a busy background, and on anything decorative. It does not read handwriting at all — that is a different class of model, not a setting. Layout is the other limit worth knowing before you start. Tesseract reads down the page, so a two-column article comes back as the whole of the left column followed by the whole of the right, rather than as you would read it. Tables lose their structure entirely: the cells arrive as separate lines with nothing to say which row or column they came from. For a document that still has its text layer — most PDFs that were exported rather than scanned — PDF to Text will give you the characters exactly as the file stores them, with no guessing involved at all. Use this when the text only exists as pixels.
Questions
Is my image uploaded anywhere?
No, and this is the tool where that matters most. The engine is what travels: Tesseract is compiled to WebAssembly and fetched from this site as a file, then it reads a picture your browser already has in memory. Nothing about the image is sent anywhere at any point, and the page makes no network request at all until you actually drop something. That is worth stating plainly here because of what people run OCR over — receipts, invoices, passports, prescriptions, contracts, screenshots of private conversations. Every free OCR site that asks you to upload one of those has taken a copy of it.
How accurate is it, really?
Very good on clean printed text, and noticeably worse on everything else. A screenshot, a PDF page rendered to an image, or a flatbed scan of a printed document is close to the ideal case. A photograph taken at an angle, a low-resolution image, text over a busy background, or a decorative typeface will all produce mistakes, and the mistakes look like ordinary words rather than obvious garbage — a misread digit in an invoice total does not announce itself. The page shows Tesseract's own mean confidence, but treat that as a hint rather than a verdict: a confident wrong answer is entirely possible. Check anything that matters against the picture.
Can it read handwriting?
No. Not badly — not at all. Handwriting recognition is a different class of model from printed-text OCR, not a setting you can turn on, and the English model here is trained on printed type. A page of handwriting will come back as empty or as nonsense. The same goes for heavily stylised display lettering and most handwriting-imitation fonts.
Why does the first image take so long?
Because the first one also downloads the engine. Tesseract plus its English language model is about 4.9 MB, and none of it is fetched when you open the page — only when you give it something to read. After that first fetch the browser has it cached, and later images pay only for the recognition itself, which is usually a second or two for a screenshot and longer for a full-page scan. If you want it available with no network at all, the offline-access page will store it for you, by name and with the size in front of you.
What happens to columns and tables?
They lose their shape. Tesseract reads down the page, so a two-column article comes back as the whole left column followed by the whole right column rather than in reading order. A table is worse: the cells arrive as separate lines with nothing recording which row or column each came from, so you get the contents without the structure. If the text is in a PDF that was exported rather than scanned, PDF to Text will give you the characters the file already stores, exactly as it stores them, with no guessing at all — that is the better tool whenever it applies.
Which languages does it handle?
English, in the Latin alphabet. Tesseract supports over a hundred languages, but each one needs its own model file of a few megabytes, and serving all of them would mean asking every visitor to think about which to download. Only English is served here. Text in another language that uses the Latin alphabet will often partly work — the letter shapes are the same — but the accuracy comes from the language model as much as the shapes, so expect mistakes on accented characters and on words English does not contain.
Is my file uploaded to a server?
No. Transmute processes everything locally in your browser using JavaScript and WebAssembly. Your files never leave your device — there is no server, no upload, no cloud processing.