Best Urdu OCR Tools in 2026, Tested on a Real Scan (اردو OCR)

The page comes from Tota Kahani, an Urdu book from 1801, in a public-domain scan. The score is the share of characters a tool got wrong, so lower is better. FastOCR scored 9.1% and imagetotext.info 9.2%; the reference text itself is off by 1% to 6%, so that is a tie.

The rest of the field, lower is better:

• OCR.space Engine 3: 19.6%; its default Engine 2: 30.4%.
• Tesseract 5.5.4 with its Urdu pack: 55.3%.
• PDF24 with Urdu picked by hand: 69.0%.
• OnlineOCR.net: 99.6%, because it has no Urdu setting and returned Latin letters.

imagetotext.info's home page says it is powered by tesseract-ocr, yet it scored far better than our own Tesseract run, so we do not know what happens inside it and do not guess. One page is a spot check, and an 1801 book is not a newspaper or a phone photo.

If you searched for i2OCR Urdu: its Urdu page describes an Urdu-optimized model, but the site asks for a CAPTCHA, so we have no score for it and make no claim about it.

Every tool that read both our Arabic and Urdu pages did worse on Urdu. FastOCR went from 2.6% on Arabic to 9.1% on Urdu, Tesseract from 20.4% to 55.3%, and PDF24 from 35.5% to 69.0%. Lower is better.

Urdu uses the Arabic letters plus some of its own. The larger change is the print style: most Arabic print is Naskh, an upright style, and most Urdu is Nastaliq, a sloping one. A tool tuned on Arabic print meets a different shape.

Four things make Nastaliq hard for a machine:

• No flat baseline. Words slope down toward the left, so line finding that expects a straight line can split or merge lines.
• Tight clusters. Letters join and stack, so it is unclear where one ends and the next begins.
• Dots decide the letter. Several letters differ only by their dots, and a missed dot gives a different, often real, word.
• Many fonts. Books and newspapers use different Nastaliq faces, and a model that has seen few of them makes more errors.

A searchable PDF puts invisible text over the scan. FastOCR writes that layer in reading order for Urdu, so a pasted line starts with its first word and Ctrl+F finds what you see. Viewers differ: Chrome's built-in viewer can scramble Latin letters and brackets inside a right-to-left line even when the file is right, so check a second viewer before blaming the file.

The Word download marks Urdu paragraphs as right-to-left, so lines run from the right in Word. We do not claim it keeps the page layout.

FastOCR reads Urdu photos, screenshots and scans and detects the language itself, so there is no menu to set. Images need no account and get no watermark. You get raw text, a one-page searchable PDF, Word and Markdown. Start on the Urdu OCR page.

A scanned PDF comes back as a searchable PDF with the text layer in reading order, once you have a free account. A finished Urdu image can be translated or cleaned up with AI Polish; neither works on PDFs. For code there is a REST API, Python SDK and MCP server.

In our test: 9.1% of the Nastaliq page's characters wrong, level with imagetotext.info (9.2%) inside the noise and well ahead of OCR.space Engine 3 (19.6%). Lower is better.

imagetotext.info exports text, Word, HTML and Markdown, and its page tells visitors they are not asked to register. The language sentence on its home page names 23 languages and Urdu is not among them.

In our test: 9.2% of the Urdu page's characters wrong against FastOCR's 9.1%, a tie inside the noise (lower is better).

Source: imagetotext.info, checked 4 November 2026. Compare: FastOCR vs imagetotext.info.

The OCR.space web tool is free without registration, with a 5 MB cap on each document. Its API page lists the free plan at 1 MB files and 3 PDF pages, with searchable PDFs marked "with watermark".

In our test: Urdu is missing from its language table, but both engines read our page. Engine 3 got 19.6% of the characters wrong and the default Engine 2 30.4%, on a public demo key. Lower is better.

Source: OCR.space API, checked 4 November 2026. Compare: FastOCR vs OCR.space.

Tesseract is free, open-source OCR that runs on your own computer, so no file leaves it. It has an Urdu data file and writes text, hOCR, PDF, TSV and ALTO.

In our test: version 5.5.4 with the Urdu pack got 55.3% of the page's characters wrong. It is a fair tool on clean English, but Nastaliq needs more work than the tools above. Lower is better.

Source: Tesseract data files, checked 4 November 2026. Compare: FastOCR vs Tesseract.

PDF24 is free with no registration, funded by advertising, and deletes files after an hour. You choose Urdu from its language menu, and the result is a searchable PDF or PDF/A.

In our test: with Urdu picked by hand it got 69.0% of the page's characters wrong. Lower is better.

Source: PDF24 OCR, checked 4 November 2026. Compare: FastOCR vs PDF24.

OnlineOCR.net lets guests convert 5 files an hour up to 15 MB. Its language menu has no Urdu, Arabic or Persian.

In our test: with its default English setting our Urdu page came back as Latin letters, 99.6% wrong. Lower is better.

Source: OnlineOCR.net, checked 4 November 2026. Compare: FastOCR vs OnlineOCR.net.

i2OCR has a page just for Urdu. It says it uses an Urdu-optimized model for printed Nastaliq and Naskh text, and that the free version processes one image per run while bulk OCR is premium. Its FAQ allows images up to 20 MB and says outputs are removed within 30 minutes. It does not name its engine, so we will not guess one.

In our test: not run. Try the same page there and on FastOCR and compare the pasted lines. Source: i2OCR Urdu, checked 4 November 2026.

Open a PDF or a .jpeg, .png or .gif photo with Google Docs to get its text. The help page lists Urdu, wants text at least 10 pixels high, and says tables and columns are not likely to be detected. We did not test it, so we give no figure.

Source: Google Drive Help, checked 4 November 2026. Compare: FastOCR vs Google Drive OCR.

iLovePDF makes a searchable PDF and lists Urdu among its document languages. Its free plan is described as limited document processing.

Source: iLovePDF OCR PDF, checked 4 November 2026. Compare: FastOCR vs iLovePDF.

NewOCR says it needs no registration, has no limit on the number of files and supports 122 languages including Urdu. Its page says the service is based on the Tesseract OCR engine, so our Tesseract run is the nearest evidence, and it scored poorly.

Source: NewOCR, checked 4 November 2026.

ABBYY FineReader PDF 16 claims 198 recognition languages. We looked for Urdu on its specifications page and did not find it, while Arabic is listed. That may be a gap in the page and not in the product, so ask before buying. We did not test it.

Source: ABBYY specifications, checked 4 November 2026. Compare: FastOCR vs ABBYY FineReader.

We did not measure these; each is cheap to try:

• Keep the page straight and fix a tilted photo before you upload.
• Use a sharp image with even light, and crop out shadows, fingers and desk.
• Check a few lines against the page, because a missed dot gives a real but wrong word.
• For a PDF, copy a line first: if it pastes as clean Urdu, the file already has text.

• A scanned Urdu book or notice as a PDF: use FastOCR for a searchable PDF; the Urdu PDF to text page has the steps.
• One photo, text only: Urdu image to text or imagetotext.info.
• Handwritten Urdu: we did not test it; see Urdu handwriting OCR and expect to proofread.
• A developer prototype: OCR.space has a free API with small limits, and FastOCR has an OCR API.
• Arabic text, not Urdu: see best Arabic OCR tools.

• Run dates: the production API produced FastOCR's output on 25 October 2026, and the rest ran on 31 October 2026. Vendor facts were checked on 4 November 2026. Re-run this test before quoting the numbers after January 2027.
• Page: one scanned page of Tota Kahani from Wikimedia Commons, scored against its Wikisource transcription.
• Score: spaces, punctuation and Arabic diacritics are stripped, then we count the share of characters wrong (lower is better). The Wikisource text and our own proofread key differ by 1% to 6%, so gaps under about 3 points are noise.
• FastOCR got a 500-pixel-wide image; the other tools got 1,280 pixels.
• Tesseract 5.5.4 ran locally with the Urdu pack. OCR.space ran on its public demo key. PDF24 ran with Urdu chosen by hand. OnlineOCR.net ran with its default English setting because it has no Urdu option.
• i2OCR is behind a CAPTCHA. Google Drive, iLovePDF, NewOCR and ABBYY were not run.
• One page is a small sample. Newspapers, phone photos and handwriting were not tested.