Why we are building our own text recognition into Rutba Office

This is a guest article by the people who make Rutba Office, a free desktop office suite for Windows. It describes work in progress: text recognition is not in the suite yet.

A scanned contract, a photographed receipt, a PDF that is really a stack of pictures: the words are there for a person to read, but not for a computer to search, copy or quote. Turning those pictures back into text is called optical character recognition, or OCR. In 2026 the quickest way to add it to a product is to rent it. You send the page to a service, the service reads it, and the text comes back, for a fee per page.

We could have done that in an afternoon. We are not going to, and this article is about why.

Why not rent it

Because the page would leave your computer. Rutba Office was built around one promise: it does not contact anything on its own. The scans people most want to search are identity papers, medical letters, bank statements, legal bundles and old family records. Those are exactly the pages they should not have to hand to a third party in order to find a word in them. If recognition needs a server, it breaks the one promise the suite is built on, so it has to run on the machine, with the network switched off.

Because a rented service decides what it reads well. Cloud OCR is very good at the languages its largest customers use. Our users write Urdu, Arabic and Hindi as well as English, and we are adding Chinese and Japanese. A service we rent will not be tuned for the newspaper Urdu a reader in Lahore wants to search, and we would have no way to make it better.

Because a service can change its price, its terms or its mind. A free suite that depends on a paid service per page is not really free, and a feature that stops working when someone else's contract changes is not really ours.

Why not simply bundle someone else's engine and stop there

Free, open recognition engines exist and run offline, and we will use one to begin with. But stopping there would leave our users with the weaknesses those engines have, and the script that shows it most clearly is Urdu.

Most printed Urdu is set in Nastaliq, a flowing style of the Arabic script in which letters stack and overlap along a diagonal. Published research finds that the open engines cope with the plainer Naskh style of Arabic script and struggle badly with Nastaliq; one study measured an open engine recovering only a small fraction of the words on Nastaliq pages. For tens of millions of people, that is the difference between OCR that works and OCR that does not.

Nobody else is going to fix that for us on our timetable. If we want Rutba Office to read Urdu the way it is actually printed, we have to build the part that reads it.

Why we think we can

Building recognition from nothing is years of work, and we are not pretending otherwise. What makes our own models practical is something the suite already has.

To learn to read a script, a model needs a very large number of lines of text together with the right answer for each. Rutba Office already draws every script it supports correctly, Nastaliq included, using freely licensed fonts. So we can render hundreds of thousands of lines of real Urdu text, age them with blur, noise, skew and the marks of cheap scanning, and teach a small model to read them, knowing the answer for every line. Researchers have already shown that a model of this kind, trained on Urdu, reads Nastaliq far better than general engines do. The model is small enough to ship inside the download and run on an ordinary laptop.

The rest of the work is already ours either way: straightening a crooked scan, finding lines and columns, working out where a table's cells are, and writing the text back into a PDF without redrawing its pages.

How we will keep ourselves honest

Owning it means we cannot blame anyone else when it gets something wrong, so we will measure it in public.

The first version starts with the open engine Tesseract, compiled to run inside the suite, with its language models in the download and nothing fetched when you first use it. Recognition runs on background threads, so a long scan never freezes the window you are working in. Alongside it we train our own models, Urdu Nastaliq first. Every model, ours or not, is tested on the same pages: pages we render ourselves in each language, so we know exactly what they say, and real scans and photographs. We count the errors per language, the time per page and the memory used on an ordinary four-core laptop, and those numbers go in the release notes. Where our model reads better, it replaces the engine for that language. Where it does not, it waits. The numbers decide, not our pride.

Where you will see it

• PDF: a scanned PDF gets an invisible text layer under each page, so Find, copy and paste work on it, while the page itself is left exactly as it was.
• Pictures and Image: Copy Text from Picture, with the recognised lines selectable over the photo.
• Worksheets: Data from Picture, which turns a photo of a printed table into cells, shown for you to check before they land.
• Mail: scanned attachments become searchable along with the rest of your mail.

The first version reads English, Urdu, Arabic, Hindi, Chinese and Japanese. As the suite learns more languages, recognition will follow them one at a time, so that every language Rutba Office speaks, it can also read.

What it will not do

It reads printed text. Handwriting is a different problem, and a bad handwriting reader is worse than none. Complicated magazine layouts and photos taken at steep angles will be read less well than clean scans, and Urdu will carry a beta label until our own model has earned it off. We would rather say that plainly now than have a review say it later.

When

The work started this week. Text recognition will arrive in a coming release of Rutba Office once it passes the same checks every release passes, with its measured accuracy in the release notes. Rutba Office is free, with no account and no paid tier, at office.rutba.io.