Sunday, Oct 18, 2026 09:45 [IST]
Last Update: Saturday, Oct 17, 2026 16:49 [IST]
PART-1
1. Introduction: The Digital Journey of the Nepali Language
Nepali is a constitutional language of India and the official language of Nepal. It is spoken by approximately 32 million to 40 million people in India, Nepal, Bhutan, the United States, Myanmar, and various diaspora communities. The language is written in the Devanagari script. In the field of computer technology, however, Nepali is still classified as a low-resource language. This does not mean that Nepali lacks speakers or literature; rather, fewer digital tools and resources are available for Nepali compared with English and certain other languages.
Over the past two decades, linguists, engineers, and various institutions have made significant contributions to building the digital foundations of the Nepali language. Alongside initiatives undertaken in Nepal, India has also implemented various programmes in this field because Nepali is one of India's 22 Scheduled Languages. Digital resources and tools for Nepali now include fonts, keyboards, dictionaries, translation systems, speech recognition, and artificial intelligence applications. Nevertheless, considerable work remains to make these resources more comprehensive, reliable, and practical.
During the past two years (2025–2026), the pace of technological development for the Nepali language has increased considerably. New open language models, speech systems, tokenizers, and datasets have emerged in both Nepal and India, and government-level cooperation between the two countries has begun. This article describes that journey, analyses the remaining technical and policy-related shortcomings, and outlines the steps that need to be taken next.
Between 2000 and 2010, efforts in Nepal primarily focused on making the Nepali language easier to use on computers. The Font Standardization Project, supported by UNESCO, developed Unicode-compatible Nepali fonts and keyboard drivers.
The PAN Localization Project, led by Madan Puraskar Pustakalaya, and the Bhasha Sanchar Project developed several language-processing tools. These included a Nepali spell-checker, thesaurus, Dobhase English-to-Nepali machine translation system, stemmer, morphological analyser, part-of-speech tagger, text-to-speech technology, and optical character recognition (OCR) technology for identifying characters in images.
These initiatives contributed to the creation of the Nepali National Corpus and the localisation of Linux for use in Nepali. The Information and Language Processing Research Laboratory and Language Technology Kendra at Kathmandu University, established in 2004, provided an institutional foundation for this work.
The use of machine learning in Nepali language technology began to expand around 2015. Google and Microsoft incorporated Nepali into their technologies, and the language became available on Google Translate.
In 2022, NepBERTa, one of the first large language models developed for Nepali, was released. The Nep-gLUE evaluation benchmark was also introduced. In 2024, researchers released pretrained BERT, RoBERTa, and GPT-2 models based on a 27.5 GB corpus. These models were considerably larger than those previously available.
In 2025, NepaliGPT further advanced the automatic generation of text in Nepali. Speech technology also continued to develop during this period, and character recognition technology was extended to the historical Nepal Script. More recently, Kathmandu University, with support from Google, developed a trilingual machine translation system for Nepali, Tamang, and English.
The pace of development accelerated further in 2026. Himalaya AI, an open research initiative based in Kathmandu, released HimalayaGPT-0.5B, together with its Nep-Tokenizer and large Devanagari pretraining corpora. It also announced plans to expand its dataset to one trillion tokens and train models with three to eight billion parameters.
Another development is Arkios, a bilingual English–Nepali model with 1.04 billion parameters, trained from scratch on 150 billion tokens and using a Devanagari-aware tokenizer. It has been released under the Apache 2.0 licence. Other open initiatives include neBrahma, a Nepali pretraining corpus containing approximately 1.85 billion tokens, and GPT-2-based Nepali models trained using custom byte-pair encoding (BPE) tokenizers.
In speech technology, Naamche Labs released Kriti, an open Nepali speech-recognition model with 119 million parameters, developed using approximately 630 hours of audio.
In August 2026, Kathmandu University's Information and Language Processing Research Laboratory (ILPRL), the Central Department of Linguistics at Tribhuvan University, and Himalaya AI Labs signed a tripartite agreement to develop benchmark datasets containing text, speech, and images for Nepali and other languages of Nepal. They also agreed to advocate for Nepal AI Sangraha, a proposed centralised open-data portal for the Government of Nepal.
As Nepali is one of India's 22 Scheduled Languages, it has been included in the Government of India's language technology programmes. Since the 1990s, the Technology Development for Indian Languages (TDIL) programme, operating under the Ministry of Electronics and Information Technology, has developed various digital resources for all 22 languages, including Nepali.
These resources include Unicode fonts and keyboards, spell-checkers, optical character recognition, text-to-speech technology, automatic speech recognition, machine translation, linguistic corpora, dictionaries, and IndoWordNet, a network of semantically related words. These resources are made available through the TDIL Data Centre.
Several institutions have collaborated in this field. For example, the GIST group at the Centre for Development of Advanced Computing (C-DAC) in Pune developed AnglaBharati, an English-to-Nepali translation system. This system is hosted on the government cloud. The group has also developed Unicode typing tools that support Nepali.
Launched in 2022, the National Language Translation Mission, Bhashini, provides a unified platform for speech recognition, machine translation, text-to-speech, and character recognition technologies for Scheduled Languages. Models for Nepali speech recognition, translation, and speech synthesis are also available on this platform. Bhashini supports the operation of more than 800 government websites.
A particularly important development for Nepal is the collaboration between Bhashini and Kathmandu University. Under this collaboration, the two sides have agreed to jointly develop Nepali datasets, speech corpora, speech-to-text systems, text-to-speech systems, and machine translation. This initiative is expected to establish a foundation for the development of voice-based technologies in Nepal.
In June 2026, the Digital India BHASHINI Division and Kathmandu University's Centre for Digital Public Infrastructure and Artificial Intelligence signed a Memorandum of Understanding (MoU) to jointly develop multilingual language technology and digital public infrastructure. The agreement covers high-quality Nepali datasets, speech corpora, speech-to-text, text-to-speech, machine translation, multilingual conversational AI, joint research, training, and pilot projects. It also includes support for digitising the linguistic and literary heritage of low-resource languages across the India–Nepal region.
Similarly, IIT Madras's AI4Bharat project released IndicTrans2, an open-source translation model covering all 22 Scheduled Indian Languages, including Nepali. The Central Institute of Indian Languages (CIIL) has also developed an important collection of Nepali text and speech corpora and dictionaries.
The achievements in Nepali language technology are not confined to research institutions. Many resources are already available for public use.
A large dataset containing approximately 157,000 sentences required for Nepali speech recognition can be downloaded from OpenSLR. High-quality Nepali text-to-speech recordings collected by Google are also available. Both resources are available under open licences. In addition, Mozilla Common Voice provides community-contributed Nepali speech data.
The Central Institute of Indian Languages provides a gold-standard Nepali text dataset and a Nepali speech dataset. The speech dataset contains 43 hours of sentences recorded from 346 speakers. These materials are available free of charge for research and educational purposes, while commercial use requires payment.
Large datasets such as IndicTrans2, IndicCorp, and Bharat Parallel Corpus Collection are openly available through platforms such as Hugging Face and GitHub. These platforms also host community-contributed Nepali datasets containing hundreds of thousands of documents, along with Nepali language models such as NepBERTa and NepaliGPT.
India's TDIL programme provides CDs containing free language tools, Nepali Unicode fonts, and Unicode typing tools. Bhashini offers open interfaces for Nepali speech recognition and translation, and its data platform has been released as open source. Nepali is also included in Meta's FLORES-200 dataset and the NLLB-200 model.
These resources provide an essential foundation for research, translation, language teaching, and the development of artificial intelligence tools for Nepali.
The list of available resources has recently expanded to include the Arkios and HimalayaGPT model weights and tokenizers, the Kriti speech model, large Nepali pretraining corpora such as neBrahma, and the Himalaya AI tokenizer corpus. Many of these resources are available under open licences on Hugging Face. Nepali evaluation datasets have also increased, including the NLUE benchmark and datasets designed for specialised tasks such as sentiment analysis and event extraction.
TO BE CONCLUDED
(Email: khagen.sharma@gmail.com)
