image to text
The reader runs in this tab — and it tells you how sure it is about what it read.
Up to 30 MB. Sharp, straight, well-lit text reads best. A photo taken at an angle is the usual reason a word comes out wrong.
Checking whether the text reader is already on this device
0%The short answer
Tesseract — the OCR engine Google open-sourced and most of the internet quietly runs on — is compiled to WebAssembly and reads the picture inside this tab. It is a 4.6 MB download the first time you use it and nothing on any visit after that. A screenshot with capitals ten pixels tall comes back word for word; a photograph of that same small print comes back as nonsense, and the confidence figure on the page drops to the teens to tell you so.
Drop in a screenshot, a photo of a page or a scan, and get the words back as text you can copy. Free, no account, no sign-in — and it shows you how confident it was, so you know which figures are worth a second look.
The formula
picture → greyscale → line and word boxes → a neural net reads each word → text
Worked examples
- A screenshot, capitals 10 px tall
- every word right
- A photo of that same small print
- nothing usable — zoom in
- A clean page held 7° off square
- one word in seven wrong
Reference table
| Capital height | Clean screenshot | Photographed or scanned |
|---|---|---|
| 23 px | 100% | 100% |
| 14 px | 100% | 86% |
| 10 px | 100% | unusable |
| 8 px | 86% | unusable |
| 7 px | 71% | unusable |
How to use it
- 01Open the page and give it a secondThe reader starts arriving in the background as soon as you land, unless your connection looks metered — then it waits until you actually pick something.
- 02Drop in a screenshot, photo or scanPNG, JPEG, WebP, GIF, BMP and TIFF all work. What matters is not the megapixel count but how many pixels tall the letters are, and whether the page is square to the camera.
- 03Read the confidence figure before you trust a numberNinety-something means copy it and move on. When we fed it print too small to read, the figure fell to the teens — it is worth believing when it says it struggled.
Also searched for
Searches that land here include image to text, ocr online, extract text from image, screenshot to text, photo to text converter and picture to text.
Getting words out of a picture
What OCR is actually doing
It is not matching shapes against a library of letters — that was OCR in the 1990s, and it is why old scanners mangled anything unusual. Tesseract turns the picture grey, finds the lines, finds the word boxes inside them, then runs a neural network along each word predicting one character at a time from what it has already seen. That is why it reads a whole word more reliably than a single letter in isolation, and why a rough idea of what English looks like helps it guess.
Two things go wrong, and they multiply
Tilt is the first, and it is the one people do not expect. On a clean page:
| Tilt | Words read correctly | | --------- | -------------------- | | Square on | 100% | | 3° | 100% | | 5° | 93% | | 7° | 86% | | 10° | 57% | | 15° | 36% |
The reader straightens a small skew internally and gives up past a few degrees. Flat paper, camera square above it, crop tight — that alone fixes most bad results.
Size is the second, and it does not act on its own. Big print shrugs off a bad photograph entirely: capitals twenty-three pixels tall read every word even after we added noise and blur. Small print has no margin left to lose — at ten pixels a clean screenshot was still perfect, while the photographed version returned nothing worth keeping. If you can only fix one thing, get closer.
What it will not do
Handwriting. Stylised logos. Text printed across a busy photograph. Anything turned 90° to the rest of the page. It also has no notion of reading order in a magazine layout: you get all the words, roughly in the order it found them, which is not always the order a person would read them in.
Questions people ask
Why did it get a word wrong?
Either the print was too small or the page was tilted, and it is worth knowing which. We rendered the same five lines of text and rotated them: square on, every word came back; at 5° it lost one word in fourteen; at 10° it lost nearly half. Separately, we shrank the text: a clean screenshot with capitals ten pixels tall still read perfectly, but the same page photographed at that size came back as nonsense. So the fix is one of two things — hold the page flat and square, or get closer so the letters are bigger. A better camera is rarely the answer.
How big does the text need to be in the picture?
It depends entirely on how clean the picture is, and the two effects multiply rather than add. On a screenshot — no noise, no blur — capitals ten pixels tall read perfectly and even seven pixels gets two words in three. Take the same small print with a camera and it collapses completely: at ten pixels we got nothing usable at all, while the same photograph of larger text, capitals twenty-three pixels tall, still read every word. The rule of thumb people repeat is to scan at 300 dpi, which puts a 10 pt capital at about twenty pixels. Our measurements say that is right for anything photographed or scanned, and unnecessarily strict for a screenshot.
Can I trust the confidence figure?
More than we expected, which is why it is on the page. Across everything we measured, the number tracked reality: 95% when every word was right, 87% when one word in seven was wrong, and 17% on the pictures where the output was gibberish. It is a per-page average, so a single mangled word inside an otherwise clean page will not drag it down much — but if it reads in the seventies or below, the reader is telling you something is genuinely wrong with the picture rather than with one word in it.
Can it read handwriting?
No, and it will not fail gracefully — it returns something confident and wrong. Tesseract was trained on printed and typeset text. Neat block capitals sometimes survive; ordinary joined-up writing does not. If you need handwriting read, this is the wrong tool, and no browser-based one will do it well today.
Which languages does it read?
English only, for now. Each language is its own two-megabyte model file, and quietly downloading five of them so that one visitor in fifty gets Portuguese is not a trade we are willing to make on everyone else's data plan. English text in any Latin-script layout is fine — columns, receipts, forms, code — and Cyrillic, Arabic or Chinese will come back as a mess rather than as an error, which is worth knowing before you paste the result anywhere.
Is my picture uploaded anywhere?
No, and there is nowhere for it to go: this site has no server that accepts files. The engine, the model and the picture all meet inside the tab, and the picture is gone when you close it. You do not have to take that on trust either — open your browser's network panel while you use it, and you will see the model come down once and nothing go up, ever.
Why is there a 4.6 MB download?
Because that is what OCR weighs. About 2.9 MB is Tesseract itself compiled to WebAssembly, and 2.0 MB is the trained English model — the file that actually knows what an 'a' looks like. Every site that reads text from pictures without that download is doing it on a server, which means your picture went there. This one arrives once, is kept by the browser, and the second visit starts instantly.