OCR — Image to Text

No. Text recognition runs entirely in your browser using WebAssembly (the tesseract.js OCR engine). The image itself never leaves your device. The only network request is a one-time download of the language pack the first time you use a given language — after that, your browser caches it and no further requests are made.

The tool ships with a curated set of common languages — English, Spanish, French, German, Portuguese, Italian, Hindi, Chinese (Simplified), Japanese, Arabic, and Russian. Pick the language that matches the text in your image before extracting for the best accuracy.

OCR accuracy depends heavily on image quality. Low resolution, blurry photos, low contrast, skewed angles, and handwriting all reduce accuracy. For best results, use a well-lit, high-resolution, straight-on photo or screenshot of printed text, and check the confidence score shown after extraction.

Yes. The output appears in an editable text box, so you can correct any recognition errors before copying or downloading it as a .txt file.

It's possible for very neat, clearly separated handwriting, but Tesseract — the OCR engine behind this tool — is trained primarily on printed and typed text, not handwriting, so accuracy on handwritten notes is noticeably lower and less predictable than on a printed page or a typed screenshot. Cursive or joined-up handwriting is especially difficult, since the engine's character segmentation step expects letters to have visible gaps between them. If you regularly need to digitize handwritten notes, expect to do more manual correction in the editable text box after extraction than you would with a printed document, receipt, or scanned form. For genuinely handwriting-focused recognition, a specialized handwriting-OCR model would perform meaningfully better than a general-purpose printed-text engine like this one.

The first time you use a given language on this tool, your browser has to download that language's trained recognition data from a CDN — typically a few megabytes — before Tesseract can start recognizing text in it. That download only happens once; the browser caches the file afterward, so every subsequent extraction using the same language starts almost immediately, with no further network activity. Switching to a different language for the first time triggers one more one-time download for that language specifically. If your first extraction feels slow, it's this initial setup cost, not a sign that anything is wrong — and it's the only network request this tool ever makes, since the image itself is processed and never uploaded.

Yes — it isn't limited to flatbed-style scans. Photos of receipts, street signs, restaurant menus, product labels, and whiteboards all work, provided the text is reasonably in focus, evenly lit, and not too skewed relative to the camera. Real-world photos tend to score lower on the confidence metric than a clean digital scan simply because they introduce variables a scanner doesn't have — shadows, glare, curved surfaces like a crumpled receipt, and perspective distortion from the shooting angle. Flattening a receipt before photographing it, avoiding direct glare from overhead lights, and holding the camera as parallel to the surface as possible all measurably improve results on this kind of everyday, non-document text.

Any common image format your browser can display works, including JPG, PNG, WebP, and BMP — the tool accepts whatever you select or drag in as a standard image file. There's no server-side upload limit to worry about since nothing is uploaded, but very large, high-resolution images do take longer to process because recognition runs on your device's own processor, and extremely large files can be slower to load into the browser in the first place. For most everyday use — screenshots, phone photos, single-page scans — file size isn't a practical concern; it only becomes relevant for unusually large, high-megapixel images.

Not in a single pass — you choose one language before extracting, and the engine uses that language's trained character and word patterns to interpret the image. If a document mixes two languages (say, English captions on a foreign-language form), select whichever language dominates the text you most need, since the engine will still often correctly recognize individual overlapping words from a similar script even when the language setting doesn't perfectly match. For a document that's genuinely and substantially bilingual with different scripts, you'd typically get better results running the extraction twice with each language selected and comparing the two outputs.