Convert WebP to TXT

Nobody exports WebP from a camera or a scanner. It is a delivery format for web pages, and essentially every WebP anyone holds arrived by saving an image off a site — which is why this conversion has a much more specific audience than the other recognition pages here.

What people save tends to be information published as a picture: a chart, a price table, a comparison, a quotation, a screenshot somebody posted, a document that was scanned and put on a page as an image. In every case the words are visible and unselectable, and reading them back is the only route short of retyping.

Web images are served at the size the page needed. A chart displayed at 600 pixels wide is stored at roughly that, often less on a page built for speed — and text recognition needs enough pixels per character to work with.

So results from web images vary far more than results from scans or phone photographs, and the cause is almost always the same. Enlarging the file afterwards does not help; it adds pixels without adding information and usually softens the letterforms, making the reading worse rather than better.

This is the single most useful thing on the page. Many sites serve several sizes of the same image and display the smallest that fits — so opening the image in its own tab, or looking at the page source for a larger source in the set, frequently finds a version two or three times the size.

Two or three times the pixels is the difference between a mangled reading and a clean one. It takes thirty seconds and it beats every setting on this page put together.

The output is plain text in reading order. For a chart that means you get the axis labels, the series names and the numbers, and you do not get any indication of which belonged to which — the arrangement that made it a chart is spatial, and plain text has no space.

That is still useful when the point is to capture the figures. It is not useful for reconstructing the chart, and for a table published as a picture, expect to spend as long rebuilding the columns as reading them. Where the underlying data exists somewhere, finding it is faster than recognising a picture of it.

Recognition matches shapes to letters without understanding anything, and it has no way to flag where it was unsure. The output is fluent and complete-looking whether or not every character was read correctly.

That matters particularly for the material saved off web pages: prices, statistics, dates, percentages, references. A misread digit in a figure you then quote is the error most likely to survive a glance and least likely to be forgiven. Check the numbers against the image.

The recogniser works inside one language’s letterforms and vocabulary, and given the wrong one it does not report a failure — it returns the closest match available in a set that cannot contain the right answer.

This comes up constantly with web images, because a page you are reading in translation, or a screenshot from another country, is not in the language your machine defaults to. If a result looks strange in a way you cannot explain, this is the first thing to check.

Recognition happens inside the browser tab on your own processor. The image is never sent anywhere, there is no account and no daily allowance, and a folder of saved pictures converts in a single pass.

It also means the conversion works on material you would not want to hand to a third party — an internal dashboard somebody screenshotted, a document shared in a private group, a page behind a login. Those are a large share of what gets saved as an image and read back.