Rohingya OCR

An OCR model that reads printed Hanifi Rohingya text and turns it into Unicode text (U+10D00–U+10D3F). Google Lens and normal PDF tools can't read this script yet.

Made by Ahkter Husin, a Rohingya developer, for the Rohingya community.

  • File: rohingya-ocr.safetensors (14 MB), a Kraken 7 text-line recognition model.
  • Version: v1.0 (November 2026)

What it reads

  • Printed Hanifi Rohingya books and documents, in the common book fonts.
  • Hanifi Rohingya digits.
  • It doesn't read handwriting. Blurry photos give weaker results.

This is an early version: it makes mistakes, so please check the text. The training data and code are not public.

How it works

 page photo or PDF page
        │
        â–¼
 1. Find the lines ... each text line, in reading order
        │
        â–¼
 2. Read each line ... right to left → Unicode text
        │              (this model)
        â–¼
 3. Spelling rules ... Sakin + Sakin → Vowel U, and
        │              Sakin after the 14 consonants
        │              that need it at a word end
        â–¼
 4. Output ........... text to copy, edit and search

This model does step 2. Steps 1, 3 and 4 are done by my own reading program, which is not public.

The model looks at the picture, not the font. Books typed in Arif Noor Muzhary (an older font that put Hanifi letters on Arabic code points, used in most books before about 2020) and books typed in Rohingya Gonwanya or Noto Sans Hanifi Rohingya (real Unicode) all come out as the same Unicode text.

See it work: Qaida page 20

A real page from a Rohingya grammar book (Qaida), a PDF page that is only a picture, with no text inside. Everything below is the real output of this model, v1.0.

1. Find the lines. Each teal box is one line the OCR found, numbered in reading order.

Qaida page 20 with a numbered box on every line the OCR found

2. Read each line. Left: the line as it is printed on the page. Right: the Unicode text the model wrote, not yet fully checked by a Rohingya reader. The drill lines with red letters and long stretched strokes (9–19) are the hardest. In the lines checked so far, most mistakes are in the small marks above and below the letters.

Every line of Qaida page 20 next to the Unicode text the model wrote

Results: model v1.0 (9 November 2026)

Test Result
Every shape alone: digits, numbers, letters, and letter + sign + vowel units, in 3 fonts 67.7% right (1,937 of 2,862)
Numbers alone, like page numbers (00–99 and 3-digit numbers) 90–93% right
Single digits 90–100% right
The rare letter Kinna Wa on Qaida page 20 22 of 24 read
Full practice lines, letters wrong: Rohingya Gonwanya / Arif Noor Muzhary / Noto 18.5% / 21.8% / 15.6%
Real library pages: words found in the Rohingya dictionary 35.2% (1,153 of 3,278)

Honest notes:

  • Single letters and numbers are read well, but full lines are still weak: about 1 letter in 5 or 6 is wrong on new practice lines. Long stretched strokes between letters are read right only about 40% of the time.
  • The real-page score counts words found in the dictionary, which is not the same as exact accuracy. One of the three test books is a newspaper scanned too small to read.
  • The real measure will be the letter error rate on pages checked by a Rohingya reader. The goal is under 2% on clean printed books.

Please support this work

This model is free, but making it is not. It takes a lot of my own time and work: collecting Rohingya books, preparing and checking training text, training the model on my own computer, and fixing its mistakes one by one.

I do this work for my community, and I am committed to continuing it. But building tools like this takes sustained time and resources, and good intentions alone cannot sustain it. Your support would allow me to give this OCR the time it needs, and to keep developing free tools for the Rohingya language.

With your support I can:

  • teach the OCR more book fonts and older, lower-quality scans,
  • put a free online OCR on rohingyahub.org that anyone can use,
  • add a "scan Hanifi text" feature to the free Rohingya apps,
  • turn more Rohingya books and documents into searchable text before they are lost.

NGOs, universities, foundations, companies and private donors are all welcome. Any amount helps.

  • Donate: go to rohingyahub.org and press the Donate button at the top of the page (PayPal, Touch 'n Go / DuitNow, Wise, Western Union and more).
  • Grants, sponsorship or partnership: contact me.

Thank you.

About me

I'm Ahkter Husin, a Rohingya software developer, interpreter and translator from Maungdaw, Arakan (Burma). I live in Johor Bahru, Malaysia. I taught myself programming, starting in 2013, and I build free tools for the Rohingya language. Full resume: rohingyahub.org/about/resume

What I've built

  • The first Hanifi Rohingya keyboard for Android, launched on 14 January 2019, soon after the script was added to Unicode 11.0 (launch video).
  • A Hanifi Rohingya keyboard for iPhone and iPad (2022), made together with Burmese developer Kyaw Zay Ya Lin Tun.
  • RohingyaHub.org, a free platform with a 4,688-word bilingual audio dictionary, the Quran in Rohingya (114 surahs with recitations), 45 Unicode fonts and a legacy-font converter.
  • islamic_hijri_calendar_rhg, an open-source Flutter package: Hijri calendar with Rohingya (Hanifi and Latin) support.
  • This OCR, which reads printed Hanifi Rohingya books and documents.
  • Rohingya songs and Tarana poems on Spotify, Apple Music, YouTube Music and Amazon Music, to keep our oral heritage alive (Spotify).

Work

  • 2023 – now: independent business owner and software engineer, Johor, Malaysia. Built and run RohingyaHub.org.
  • 2020 – 2022: community interpreter and translator, Johor Outreach Community Center (JOCC), Malaysia. Interpreted for Rohingya refugees in UNHCR registration, health care and legal hearings, and translated documents between Rohingya, Burmese, Malay and English.
  • 2018 – 2019: general worker in a factory (Winsen Industries, Pekan Nanas, Johor) after arriving in Malaysia as a refugee, coding at night.
  • 2017: household demographic surveyor and interpreter, Médecins Sans Frontières (MSF), Cox's Bazar, Bangladesh. Interpreted between MSF doctors and newly arrived Rohingya refugees.
  • 2013 – 2016: commercial accountant, Adila Enterprise, Bangladesh.

Education

  • 2012: B.Sc. Chemistry at Sittwe University, Arakan. I could study for only a few days before the 2012 violence forced me out.
  • 2011 – 2012: Class 10 (matriculation) with distinction in Biology, Nga Khu Ya High School, Maungdaw.
  • 2005: primary school, Ahtet Phyuma Primary School, Maungdaw.

Languages

Rohingya (native: Hanifi, Latin and Arabic scripts), Rakhine, Burmese, English, Bangla, Malay, Urdu and Hindi.

I'm available for translation, interpretation, software work and Rohingya language data: contact me.

Licence

CC BY-NC 4.0. You may use and share the model for free, but not sell it or use it in a paid product. Please credit:

Rohingya OCR model by Ahkter Husin

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support