India is charting a bold new course in artificial intelligence, one that speaks directly to its people—literally. The IndiaAI Mission, a government-backed initiative launched in 2024, is now prioritising the development of voice-first, vernacular large language models (LLMs) to serve the co...
India is charting a bold new course in artificial intelligence, one that speaks directly to its people—literally. The IndiaAI Mission, a government-backed initiative launched in 2024, is now prioritising the development of voice-first, vernacular large language models (LLMs) to serve the country’s vast and linguistically diverse population. This strategic pivot, confirmed by IndiaAI Mission CEO Abhishek Singh, marks a significant shift toward inclusive, culturally rooted AI that can understand and respond in the languages of everyday Indians.
Key highlights from the initiative
1. IndiaAI Mission is investing in foundational LLMs that prioritise voice interaction over text
2. The models will be trained on regional languages and dialects, with a focus on natural speech and emotional nuance
3. Startups like Sarvam AI, Gnani.ai, and Soket AI Labs have been selected to build sovereign, multimodal models
4. The mission includes open-sourcing these models to accelerate adoption and innovation across sectors
5. Over 10,000 GPUs have been allocated to support compute-intensive training for these models
Why voice-first matters in India
- With over 900 million internet users and varying literacy levels, voice interfaces offer a more accessible entry point to digital services
- Voice commands are easier to use on mobile devices, especially in rural and semi-urban regions
- Regional voice data captures accents, tones, and informal speech patterns that text-based models often miss
- The voice commerce market in India is projected to grow from USD 1.57 billion in 2024 to USD 7.47 billion by 2030
The linguistic challenge
India’s linguistic landscape includes 121 languages and over 19,500 dialects. Traditional AI models, trained primarily on English or standard Hindi, struggle to capture the nuances of regional speech. This leads to stiff, outdated responses that feel foreign to users. The new voice-first LLMs aim to change that by training on real, expressive speech data collected from across the country.
Startups are getting creative to build this data
- AiVanta is working with over 1,200 voice artists and creators to record natural speech in eight Indian languages
- Rian.io is training voice-over artists in emotional modulation across Hindi, Marathi, and English for AI-powered translation and localisation
- Quansys AI is crowdsourcing voice samples from everyday speakers via WhatsApp, building a grassroots dataset that reflects real-life usage
Government support and strategic goals
- The IndiaAI Mission has earmarked ₹1.2 billion for foundational model development
- Sarvam AI received the largest single subsidy of ₹98.68 crore to access 4,096 Nvidia H100 GPUs
- The models will be released under permissive open-source licenses to foster public-interest use cases
- The initiative aims to reduce dependency on foreign AI models and build sovereign infrastructure for India’s digital future
Applications across sectors
- Education: Voice-enabled tutoring in regional languages for students with limited literacy
- Healthcare: AI assistants that understand patient queries in local dialects
- Governance: Multilingual chatbots for citizen services and grievance redressal
- Commerce: Voice-based shopping and banking interfaces tailored to regional users
The road ahead
India’s voice-first LLM strategy is not just about technology—it’s about identity, inclusion, and empowerment. By building models that understand the emotional and cultural depth of Indian languages, the country is laying the foundation for AI that feels familiar, not foreign. As these models go live, they promise to transform how Indians interact with machines—making digital services more intuitive, accessible, and human.
Source: Outlook Business, Forbes India, YourStory, IndiaAI.gov.in, MSN