Convert JPG to TXT

Optical character recognition looks at shapes and decides which letters they are. It does not know what the document says, and it has no way to tell you how confident it was — so the output is fluent, complete-looking text that may contain errors nobody flagged.

That is the single most important thing to understand before relying on a result. Read it against the original, especially the numbers: a misread digit in a price, an account number or a date is exactly the error that survives a quick glance and causes a real problem later.

More than any setting, more than any tool. Straight, well-lit, in-focus printed type at a reasonable size comes out close to perfect. Everything that degrades the picture degrades the reading, in roughly the order you would expect: focus first, then shadow across the text, then the angle of the page, then the size of the type in the frame.

Which means the useful lever is the camera, not the converter. Hold the page flat, shoot straight down rather than from a chair, get light from the side rather than behind you, fill the frame with the page, and tap to focus before pressing the button. Ten seconds of that is worth more than any amount of processing afterwards.

This is the setting that matters and the one people skip. The recogniser matches shapes against the letterforms and vocabulary of a particular language, and given the wrong one it does not fail — it produces confident nonsense, because it is finding the best match within a set that does not contain the right answer.

A German page read as English comes back with umlauts mangled and words that nearly exist. A document with a language it was not told about produces the same. If a result looks strange in a way you cannot explain, check this before anything else.

The output is plain text: no headings, no columns, no tables, no positions. A single-column letter comes back readable and in order. A two-column page can come back with the columns interleaved, because the recogniser reads across lines that look like lines.

A form is the hardest case — labels and values end up adjacent in ways that are hard to reconstruct, and a table becomes a run of numbers with no idea which column they came from. For those, the honest advice is to expect to do some work on the result, or to photograph one field at a time.

The recognition here is built for printed type, where a letterform is one of a few thousand possibilities. Handwriting is not: every writer produces different shapes, and the same writer produces different shapes on different days.

Neat, separated block capitals sometimes come out usable. Ordinary joined-up writing generally does not, and the output will look like real words rather than like an error — which is worse than failing, because it invites belief. If a handwritten note has to become text, typing it is faster than correcting a recognition of it.

Receipts, for expense claims. Letters and forms that arrived on paper. Pages from books and journals for a citation. Whiteboards after a meeting. Serial numbers and labels on equipment. Menus, signs and notices while travelling.

Almost all of it is personal or commercial: names, addresses, amounts, account details, medical information. That is the context in which the next section matters, and it is why it is on the page rather than in a footer.

The recognition runs inside the browser tab on your own processor, on your own device. The photograph is never uploaded, so there is no server copy, no retention period to ask about, no account and no daily allowance.

Watching the network tab while it works is the honest way to confirm it: no request carrying the image goes out. For a picture of a prescription, a bank letter or a signed contract, that is the substantive difference from a hosted service rather than a feature line.