Hindko Language AI: CECOS Builds OCR and Tokenizer Tools

Peshawar’s CECOS University is trying to bring Hindko language AI out of the research lab. It is building an optical character recognition (OCR) model that can read Hindko text, along with a dedicated tokenizer meant to be the base of a future Hindko language model. The work runs through the university’s Center for Digital Governance and Agentic Innovation (CDGAI) together with the Khyber Pakhtunkhwa Chief Minister’s Policy Office (CMPO). It is a first step toward giving one of Pakistan’s larger regional languages the kind of digital tools Urdu and English speakers already use every day.

Dawn reported the initiative on 9 November. The university’s own write-up of its October developer conference describes the technical pieces. For Pakistan’s tech sector, the project tests whether local institutions can build AI for low-resource languages themselves, rather than waiting for global platforms to get to them.

What CECOS is building for Hindko language AI

CECOS described the Hindko work after the second edition of the KP Languages Developer Conference (KPLDCon), held on 7 October at the Serena Hotel in Peshawar. CDGAI organised the event with CMPO. Linguists, software developers, researchers, language organisations and students attended.

The university named two Hindko tools. The first is an OCR model trained specifically to read Hindko, so that printed and written material in the language can be turned into machine-readable text. The second is a tokenizer built for Hindko. A tokenizer breaks text into the small units a language model actually processes. General-purpose tokenizers are trained mostly on high-resource languages, and they often cut unfamiliar words and scripts into many small pieces. That makes models more expensive to run and less accurate. CECOS says its Hindko tokenizer is meant to be the starting point for a Hindko language model.

The organisers said this year’s conference moved past the question of why local languages are missing from AI and onto the practical question of how to include them. They ended the event with a short message: “Our languages. Our data. Built together.”

A wider push across Khyber Pakhtunkhwa’s languages

Hindko was one of several languages on the conference agenda. Other sessions covered Detasawy, a community portal for Pashto datasets; bringing Seraiki into Pakistan’s AI future; and Verto 2.0, which is now adding Torwali alongside Khowar. Speakers also discussed the shared history of Hindko and Pashto, the languages of ancient Gandhara, and indigenous language communities in northern Pakistan.

The speakers included Engr. Waqar Khan, Special Secretary to the Chief Minister of Khyber Pakhtunkhwa, and Engr. Sohaib Tanveer, Vice President of CECOS University. CDGAI’s official page names Dr. Maryam Mahsal Khan as its director. It describes the center as a hub that uses AI, agentic systems and data analytics for governance, including custom language models for provincial needs and AI chatbots.

The center’s list of upcoming projects includes another language track: preserving endangered languages of Kalash and Chitral using natural language processing, speech-to-text and conversational AI. It lists low-resource language modelling with data augmentation as a sub-track.

How many people speak Hindko

Hindko is not a small language. Gallup Pakistan’s analysis of Census 2023 data on mother tongues puts Hindko speakers at about 2 percent of the national population and 9 percent of Khyber Pakhtunkhwa’s. The Gandhara Hindko Board in Peshawar, which has long promoted the language, cites the census as making Hindko Pakistan’s sixth-largest language and the second-largest in Khyber Pakhtunkhwa. Speakers also live in Azad Kashmir. The board’s sister body, the Gandhara Hindko Academy, has written a BS Hindko curriculum for universities.

The digital resources for the language are still thin. One public building block came out in February 2025: a voice dataset published in the journal Data in Brief, with 17,597 recordings of the Hindko numbers one to twenty. Students, staff and faculty at the Pak-Austria Fachhochschule Institute of Applied Science and Technology contributed the recordings.

Why low-resource language AI matters for Pakistan’s tech sector

Pakistan’s digital services are moving to chatbots, voice interfaces and automated document processing. Citizens who mainly speak a regional language risk being left out of that shift. If a model cannot read or understand Hindko, Pashto or Seraiki, it cannot answer a farmer’s question, process a local land record or digitise a century of regional literature. Tools like OCR models and language-specific tokenizers are the plumbing for all of those uses.

There is a commercial side as well. Well-documented local datasets and models give Pakistani universities and startups something global AI companies do not have. They also reduce dependence on foreign systems that handle the country’s languages poorly. Because CECOS is working with a provincial policy office, its tools could end up in government services rather than staying in academic papers.

A lot still has to happen. On the pages ITInsider reviewed, CECOS has not published how much Hindko text it has collected, how big the planned language model will be, or when it will be released. A tokenizer is a foundation, not a finished product. How useful a Hindko model becomes will depend on how much text and speech the community contributes.

Key facts

• Who: CECOS University’s Center for Digital Governance and Agentic Innovation (CDGAI), working with the KP Chief Minister’s Policy Office
• What: A Hindko OCR model and a Hindko tokenizer meant to be the base of a Hindko language model
• Where it was presented: KP Languages Developer Conference, Edition 2, Serena Hotel Peshawar, 7 October
• Other languages in the programme: Pashto, Seraiki, Torwali, Khowar; endangered languages of Kalash and Chitral are in CDGAI’s pipeline
• Speaker base: about 2 percent of Pakistanis and 9 percent of Khyber Pakhtunkhwa’s population speak Hindko as their mother tongue (Census 2023, per Gallup Pakistan)