How the OCR works in your browser
The tool uses tesseract.js, the WebAssembly version of the open-source Tesseract engine (Apache-2.0 licence), with its LSTM neural-network models for English and Arabic from the project's "best" training set. When you open a file, the engine and the language model are downloaded once from the public jsDelivr network, at fixed versions: about 4.4 MB for English, 3.1 MB for Arabic and 6.0 MB for both. The browser keeps them, so the next picture starts at once. Your picture is drawn on a canvas in this tab, turned grey, and handed to a background worker that runs the engine; only the model files travel over the network, never the image or the text. In our browser test we logged every request the page made while it read our test pictures and PDFs: none carried image data.
Accuracy we measured
On 22 October 2026 we ran the tool on test images whose exact text we knew, and compared the result character by character (character accuracy = 100% minus the share of characters that had to be added, removed or changed). The images were made on a computer: text rendered in common fonts, then turned into a scan or a phone photo with tilt, blur, JPEG compression and uneven light. Results: clean English letter scan 100%; photo of a printed English page 99.6%; phone photo of a supermarket receipt 98.5% (93.4% of words exactly right; the dashed separator lines were not counted); low-resolution screenshot with 13-pixel text 100%; Arabic printed letter 98.3% (91.8% of words); photo of the same Arabic letter 99.0% (93.9% of words); Arabic in a decorative Naskh font 97.0% (83.7% of words); a mixed Arabic and English invoice in Arabic + English mode 95.2%. Real photos of crumpled paper, small print or handwriting will do worse, so always read the result before you use it.
Arabic text: what to expect
Arabic is read correctly in most places, but it is clearly weaker than English: across our three Arabic test images about one word in ten needed a correction (one in six in a decorative font), against about one English word in fifty on a similar photo. Most Arabic mistakes are letters that differ only by their dots, such as ح and خ, ت and ن, or ب and ي, and the Arabic comma ، often comes out as an English comma. Choose Arabic alone for a page that is all Arabic: the Arabic + English mode scored 95.3% on the same Arabic letter, against 98.3% for Arabic alone, because it sometimes takes an Arabic word for an English one. On a page that really mixes the two languages, Arabic + English is the right choice: Arabic alone scored only 66.2% on our mixed invoice. Lines that start with Arabic are shown right to left, with English words and numbers kept in their order.
Getting a better result
Most errors come from the picture, not the engine. Photograph the page flat, in daylight, filling the frame, and hold the phone parallel to the paper. The tool already straightens pictures tilted by up to ten degrees and enlarges small images such as screenshots up to three times before reading them, which is why tiny 13-pixel screen text still came out perfectly. If a shadow from your hand or phone falls across the page, tick "Black and white clean-up": on our test page with a hard shadow, accuracy went from 72.9% to 100% with it. Leave it off for normal pictures, where it can lower the score slightly (98.5% to 95.2% on the receipt). If you only need one part of a page, crop it first with Crop Image so the engine is not distracted by the rest.
Scanned PDFs, speed and limits
A scanned PDF is a set of page pictures, so each page is drawn at about 215 dpi and read like a photo; type a page range to read only some pages, up to 50 at a time. On our 4-core test computer, after the engine was ready, a photo took 1 to 2.5 seconds, one scanned page about 2.5 seconds and five pages 6 to 8 seconds, because two pages are read at once on computers with four or more cores; phones are slower. The first run also needs the one-time download. The result is plain text: columns are read as separate blocks, tables become lines of words, and fonts, colours and pictures are not kept. If the PDF already has real text, PDF to Text gets it out exactly and faster. For an editable Word file use JPG to Word, and for a password-protected PDF remove the password first with Unlock PDF.