Image to text (OCR)

What OCR reads well — and what it never will

Character recognition is pattern matching, not understanding, so the result tracks the input with brutal honesty. A crisp screenshot or a flat, evenly lit scan of printed text comes out nearly perfect. A phone photo of the same page taken at a slight angle, with a shadow across one corner, drops noticeably — and handwriting, decorative fonts, text over photographs and low-contrast prints (grey on white, colour on colour) can come out barely usable. Three habits fix most of it: shoot straight-on rather than at an angle, fill the frame with the text and crop away the surroundings, and when the source is already on screen, use the original screenshot instead of photographing your monitor. Every one of those steps removes a distortion the recogniser would otherwise have to guess through.

The one-time download, and why it exists

The first time you press Read text, your browser fetches the recognition engine and its English language data — about 7 MB, served from this site and cached, so later runs start in moments. That download is the price of a real promise: the engine runs entirely on your own device as WebAssembly, which means the photo of your contract, prescription or payslip is never transmitted anywhere. Most online OCR services work the other way around — a tiny page, and your document uploaded to their server. If the first run seems to sit still for a few seconds on a slow connection, it is fetching, not stuck; the progress line tells you which phase it is in. After that first visit the engine loads from cache, and recognition on a normal page of text takes a handful of seconds even on a modest laptop.

English only, for now — and always proofread the numbers

This first version reads English text and numbers. Other languages are planned — each one is an additional trained data file, so they will arrive as separate downloads rather than one giant bundle — but today a Dutch letter or a Turkish invoice will come out garbled wherever the words stop looking like English. Numbers deserve a special warning of their own: a misread 8 for a 3 in an IBAN, a phone number or an invoice total is far more costly than a misread word, and digits carry no spelling context the engine could use to correct itself. Whatever the text, read the result once before you rely on it; whatever the numbers, check them twice.

The text is out — now put it somewhere useful

Copy text puts the result on your clipboard, and from there it usually has one of three destinations. Straight into a document or e-mail is the obvious one — the entire point of OCR is never retyping a page again. Less obvious: paste it into /text-to-speech and the page becomes audio, which turns a photographed article into something you can listen to while doing anything else. And when what you actually need is the page as a shareable document rather than loose text, /image-to-pdf wraps the original photo into a proper PDF — tidier to file and to forward than a bare image, with this tool covering the moments you need its words as well.

Getting the best result from OCR

Character recognition works best on sharp, well-lit images with dark text on a light background. Photograph documents straight-on rather than at an angle, crop to just the text if you can, and prefer the original screenshot over a photo of a screen. Currently reads English and numbers; more languages are planned.