How to Correct Urdu OCR Mistakes: What Goes Wrong and a Method That Works

Published by Shoaib Ahmed · November 2026

Image-to-text gives you text you can edit. It does not give you text you can trust. That matters most with Urdu, because one sentence can come out almost perfect from one picture and badly damaged from another. Here is a real example from our own tests: the same line, read by the Urdu OCR tool from two clean test pictures.

	Result
What was written	السلام علیکم! میں آپ کو اپنی درخواست بھیج رہا ہوں۔
Naskh-style picture	السلام علیکم! میں آپ کو اپنی درخواست بھیچ رہا ہوں۔
Nastaliq-style picture	السلام علیکم ! می ںآ پکو اپی درخواست گی را ہوں۔

The Naskh picture lost one letter: بھیج was read as بھیچ. The Nastaliq picture came out with wrong spaces, dropped letters and a changed letter. The tool itself said as much. Its confidence note read 89% (“Good”) for the Naskh picture and 66% (“Medium”) for the Nastaliq one. This article explains what goes wrong and a method for correcting it without starting from scratch.

What we tested

We took six lines of Urdu (60 words), drew them by computer in a Naskh and a Nastaliq font, and made worse copies: uneven light, a 3° tilt, a small size and camera-like noise. Computer-made pictures are cleaner than most real photos, so treat the numbers below as a best case. They come from the same test whose results are on the OCR tool page.

What actually goes wrong

Across the five Nastaliq pictures, read with the standard model:

Letters go missing more than they change. About three in four of the letter mistakes (75%) were letters that were dropped. The rest were extra letters (13%) and wrong letters (12%). Real examples: اپنی became اپی, موسم became موس, چار became جار.
Spaces do a lot of damage. 64% of the words had all their letters right, but only 46% came out fully correct, because the spaces were in the wrong place. That is almost one word in five spoiled by spacing alone: می ںآ پکو for میں آپ کو.
The first word of a line can be clipped. In one line, پاکستان was read as سان.
The full stop can be read as a dash. ۔ became - at the end of a line in three pictures. The tool’s Tidy option now turns a dash at the end of a line back into ۔.

Across the five Naskh pictures the list was short: 12 mistakes in 1,080 letters. One word, بھیج, was read with چ in all five pictures, and the final ہ and the mark in براہِ were lost in three of them.

Step 1: decide whether to correct or retype

Read the first three lines against the picture before you do anything else. If a line needs only one or two small fixes, as in the Naskh example, correct it. If most lines have wrong spaces and missing letters, as in the Nastaliq example, retyping in the Urdu online typing tool is usually faster than repairing every word, because each repair first needs you to work out what the word should have been. This is a rule of thumb, not a measured figure. In our Nastaliq test about half of the words needed work.

Step 2: fix one kind of mistake at a time

Correct spaces first, because a wrong space can hide whether a word is right. Then fix letters, then marks and dots, then punctuation and numbers. Here is the fourth line of the Nastaliq test, fixed in two passes:

Stage	Line
As the tool read it	علم حاص لکرنا ہ رملمان مرد او رعورت پر ذض ے۔
After fixing the spaces	علم حاصل کرنا ہر ملمان مرد اور عورت پر ذض ے۔
After fixing the letters	علم حاصل کرنا ہر مسلمان مرد اور عورت پر فرض ہے۔

Going in layers is faster for a practical reason. Once the spaces are right, most of the remaining mistakes show up as “not a word” (here ملمان and ذض), so you can find them by eye instead of comparing every letter with the picture.

Step 3: check what OCR cannot be trusted with

Names, numbers, dates and amounts should be checked against the picture one character at a time. The confidence note is an average for the whole page, so it cannot tell you which digit is wrong. Also look at the first and last word of every line, where clipping happens, and at marks: in our test براہِ lost its final letter and mark and became برا.

Step 4: let the tools do the repetitive work

The OCR tool’s Tidy the result option already removes extra spaces, changes ? , . into ؟ ، ۔ after Urdu text, and fixes Arabic look-alike letters. For anything that is left, such as mixed digits or repeated punctuation, paste the text into the Urdu text cleaner and press Check my text. Finally, listen to the result with text to speech. A dropped letter or word is much easier to hear than to see.

A better picture saves more time than any correction

Two settings made a measurable difference in our tests. With Improve the picture switched on, a Nastaliq picture with uneven light had 10% of its letters wrong. With it switched off, 44%. And High accuracy cut the errors on a noisy Nastaliq picture from 38% to 17%. Beyond that, crop the picture to the text, shoot straight from above the page and use even light. None of this changes the style of the print, which is why Nastaliq stays harder than Naskh.

When OCR is the wrong tool

Handwriting, very small print, decorative fonts and curved pages are all poor material for OCR. We did not test handwriting, and the tool’s own page says it cannot read it reliably. For these, typing the text is the honest answer. Our guides on learning Urdu typing and improving Urdu typing speed help when you have to do that often.

Questions people ask

Why is Nastaliq read worse than Naskh?

In our tests the same sentence came out with one wrong letter in Naskh and with about half of its words damaged in Nastaliq. A likely reason is that Nastaliq letters slope and overlap, which makes it hard for the engine to see where a letter or a word ends, but we measured the difference, not the cause.

Can I trust the confidence note?

It matched what we saw: 89% for the clean Naskh picture and 66% for the clean Nastaliq one. It is still only an average estimate, so use it to decide whether to retype, not to skip proofreading.

Is my picture uploaded?

No. The reading is done in your browser. Only the recognition engine and its Urdu language data are downloaded.