Policy Update
Ashi Verma
Background
India’s digital inclusion is constrained by its linguistic diversity, with over 90% of the population non-proficient in English, limiting access to the predominantly English internet. To address this structural barrier, the Government of India has initiated AI-driven language technologies.
BHASHINI (BHASHA Interface for India, 2022), launched by MeitY under the National Language Translation Mission, aims to develop a National Public Digital Platform for Indian languages. It seeks to provide open-source AI tools for translation, Automatic Speech Recognition (ASR), Text-to-Speech (TTS), and Optical Character Recognition (OCR) across 22 scheduled languages. Operating on a crowdsourcing model through Bhasha Daan, it enables startups and government departments to build voice first, multilingual interfaces, thereby democratizing digital services.
Building upon this foundation, BharatGen (2024-25), spearheaded by the Department of Science and Technology, is India’s first sovereign, multimodal Large Language Model initiative. Unlike global LLMs, it is trained on India-centric, balanced datasets to generate culturally contextual content in Indian languages. Together, they transition India from a consumer to a creator of linguistically inclusive AI.
Functioning
A. BHASHINI: Three-Tiered Public Digital Infrastructure Model
BHASHINI operates not as an end-user application, but as a foundational AI infrastructure based on a three layered architecture.
- Data Layer – Bhasha Daan: The primary challenge for Indian languages is lack of datasets. BHASHINI addresses this through a participatory crowdsourcing initiative, Bhasha Daan, where citizens voluntarily contribute speech, text, and parallel translations. This creates open, large-scale datasets for low-resource languages.
- Model Layer – Core Language Technologies: Using this data, national research institutions (CDAC, IIT Madras) develop core models including Machine Translation (Anuvaad), Automatic Speech Recognition (ASR), Text-to-Speech (TTS), and Optical Character Recognition (OCR) for the 22 Scheduled Languages.
- Platform Layer – Open Ecosystem: These models are deployed on the BHASHINI platform as open APIs and SDKs. This allows government departments (UMANG, e-Courts, Digital India apps) and private startups to integrate multilingual capabilities into their existing products without incurring high R&D costs, promoting innovation at scale.
B. BharatGen: Sovereign Generative AI Framework
- Indigenous Data Curation: Unlike global LLMs trained on Western-centric data, BharatGen focuses on curating balanced, de-biased, and culturally contextual Indian datasets encompassing literature, oral traditions, and domain-specific knowledge.
- Multimodal Foundation Model Training: It develops sovereign Large Language Models capable of multimodal generation text, speech, and vision natively in Indian languages, ensuring contextual and cultural fidelity.
- Sectoral Fine Tuning and Deployment: The base models are fine-tuned for sector-specific applications such as agriculture, education, and governance, enabling deployment of accurate, locally relevant AI solutions while ensuring data sovereignty.
Performance
The performance of BHASHINI since its launch in 2022 has been significant in terms of scale and adoption. Technologically, it has achieved coverage of all 22 Scheduled Languages and has notably improved translation and speech recognition accuracy for low-resource languages such as Maithili, Bhojpuri, and tribal dialects, where global platforms remain ineffective. Institutionally, its open API ecosystem has been integrated into over 50 major government platforms like UMANG, e-Shram, and e-Courts, enabling citizens to access governance in their mother tongue (Digital India BHASHINI Division, MeitY, 2026).
Furthermore, through its crowdsourcing initiative Bhasha Daan, it has mobilized lakhs of citizens to create India’s largest open-source linguistic dataset, transforming language data into a public digital good.
In comparison, BharatGen, initiated in 2024-25, is in its nascent but promising stage as India’s first sovereign multimodal LLM. Its initial Param series models have demonstrated superior cultural and contextual understanding in Indian languages compared to Western LLMs, which often produce biased or irrelevant outputs. Its sectoral pilots in agriculture and education have shown improved regional accuracy, indicating its potential to provide locally relevant generative AI solutions.
Impact
The combined impact of BHASHINI and BharatGen extends beyond technology, contributing to social, economic, and governance transformation under Digital India. Socially, they have been instrumental in bridging the linguistic digital divide by enabling over 90% of non-English speaking Indians to access digital services in their mother tongue. This has particularly empowered rural populations, farmers, and women through voice-based access to welfare schemes, legal aid, and education, while also contributing to the digital preservation of low-resource and tribal languages.
Economically, the initiatives have fostered a vibrant vernacular innovation ecosystem. By providing open source language APIs, BHASHINI has significantly lowered the entry barrier for Indian startups, promoting the development of local language products and boosting voice commerce in Tier-II and Tier-III cities.
BharatGen further strengthens this by ensuring data sovereignty and reducing dependency on foreign, biased AI models, thus paving the way for an Atmanirbhar AI economy.
In terms of governance, they represent a paradigm shift towards inclusive e-governance.
The integration of BHASHINI into platforms like UMANG and e-Courts has enabled last-mile delivery of governance. Together, they establish a UPI-like model for AI – a public digital infrastructure where the government provides the foundational platform and the private sector builds innovative solutions on top of it, truly realizing the vision of AI for All.
However, both initiatives still face constraints related to dialectal diversity, real-time performance, and high computational requirements. Overall, their combined performance marks India’s transition from merely accessing digital content to creating indigenous, linguistically inclusive AI.
Emerging Issues
- Challenges Pertaining to Data and Linguistic Diversity : A critical constraint is the uneven availability of linguistic data. While resource-rich languages like Hindi possess adequate datasets, several low-resource, tribal, and North Eastern languages continue to suffer from data scarcity. Additionally, data obtained through participatory platforms such as Bhasha Daan often lacks rigorous quality checks, resulting in non-standardized and error-prone corpora. Such datasets fail to capture India’s complex sociolinguistic practices, including code-mixing, dialectal variations, and informal speech patterns, thereby limiting model accuracy.
- Technological and Infrastructural Limitations : The operational efficiency of language models declines considerably outside controlled environments. Speech recognition systems exhibit lower precision in the presence of background noise and diverse regional accents. Furthermore, the development of sovereign multimodal models under BharatGen is contingent upon access to high-end computational infrastructure, which remains costly and largely import-dependent. The integration of text, speech, and visual modalities across 22 languages is still at a nascent stage and requires substantial research advancement.
- Ecosystem Development and Financial Sustainability : Despite its open-access architecture, BHASHINI has witnessed modest adoption within the private sector, as many developers continue to prefer established global APIs offering better performance metrics.
The shortage of specialized human resources in Indian language AI research further impedes innovation. Moreover, the absence of a self-sustaining economic model raises concerns regarding the long term maintenance, upgradation, and hosting of these public digital platforms.
- Ethical and Governance Implications : Largescale collection of voice and language data introduces significant concerns related to privacy, consent, and potential misuse. The deployment of generative models also poses risks of misinformation, cultural misrepresentation, and algorithmic bias, especially in sensitive regional contexts.
Lastly, technological intervention alone cannot guarantee digital inclusion, as structural barriers such as limited digital literacy, inadequate internet connectivity, and low smartphone penetration continue to exclude the intended beneficiaries.
Way Forward
The future roadmap for BHASHINI and BharatGen must focus on making them more accurate, inclusive, and sustainable. There is a need to improve data quality by implementing strict validation for crowdsourced contributions and launching special drives for tribal and low-resource languages in collaboration with local institutions.
Technologically, India must invest in indigenous computer infrastructure under IndiaAI Mission and develop lightweight models that work efficiently in low-connectivity rural areas.
To increase adoption, startups should be incentivized to build on these open platforms and their integration should be made mandatory in all government citizen services. Simultaneously, capacity building through specialized courses in Indian language AI is required.
Finally, a clear ethical framework ensuring data privacy, consent, and auditing for bias must be established, coupled with grassroots digital literacy initiatives, so that these initiatives truly achieve AI for all and empower every citizen in their own languages
References
Ministry of Electronics and Information Technology. (2024). Digital India BHASHINI National Language Translation Mission (NLTM).
Press Information Bureau. (2026, February 17). Minister lauds India’s first government owned Sovereign LLM, BharatGen: Supported by ₹235 Crore under NM-ICPS and ₹1058 Crore India AI Mission. https://pib.gov.in/PressReleasePage.aspx?PRID=2103202
Ministry of Information & Broadcasting. (2024). Bhasha Daan initiative
https://www.india.gov.in/spotlight/bhasha-daan-contribute-indian-languages
BharatGen & IIT Bombay. (2026). BharatGen: India’s Sovereign Multilingual AI Stack – Param, Shrutam, Sooktam, Patram models in 22 languages. https://bharatgen.ai/
Digital India Corporation. (2026, March 27). Bhashini Samudaye and Shrutlekh: Real-time transcription across 22 languages.
https://pib.gov.in/PressReleasePage.aspx?PRID=2115731
NITI Aayog. (2018). National Strategy for Artificial Intelligence. https://www.niti.gov.in/sites/default/files/2023-02/NationalStrategy-for-AI-Discussion-Paper.pdf
IndiaAI Mission, MeitY. (2024). IndiaAI Mission: Building sovereign AI ecosystem and compute infrastructure.
About the Contributor
Ashi is pursuing B.A Program (History and Political Science) from Maitreyi College (NCWEB) South Campus Delhi University with a major in Political Science. She is passionate about governance, public policy, social issues and digital innovation, with a keen interest in research, academic writing, and data documentation. Through research and policy analysis, she aims to contribute to evidence-based policymaking and create meaningful social impact.
Acknowledgement
I would like to express my sincere gratitude to IMPRI (Impact and Policy Research Institute) for giving me the opportunity to work as an intern and write on such an important topic. I would also like to acknowledge the official MeitY and IndiaAI Mission for making authentic data available on the PIB Portal that helped me to complete this article.
Disclaimer
The views and opinions expressed in this article are those of the author and do not necessarily reflect the official policy or position of IMPRI (Impact and Policy Research Institute) or any other affiliated organisation. The information presented in this article is based on publicly available government sources and is intended solely for academic and research purposes.
