India to focus on voice-first vernacular LLMs: AI Mission CEO

Mint

17 hours ago

New Delhi: India's $1.2-billion AI Mission is preparing to make voice-first artificial intelligence (AI) models for Indian languages its calling card, aiming to differentiate itself from Big Tech firms that focus primarily on text-based AI.
The differentiator for artificial intelligence from India could include Indic languages-focused large language models, alongside voice-first models.
In an interview with Mint, Abhishek Singh, additional secretary at the Ministry of Electronics and IT (Meity) and chief executive of the Centre's India AI Mission, said that the country's push to build its own foundational AI models will seek to create solutions that can be replicated in other nations.
'India's common compute infrastructure is a unique model that is receiving a lot of interest from around the world, especially the global south," Singh said. 'While US tech firms are largely building foundational models based on text data as the primary medium, as India, we see a sizeable opportunity to develop foundational models that are primarily trained on voice data, because in future, voice will be the primary and most natural way in which people will interact with AI."
The AI Mission, announced in March last year, seeks to offer funding support in the form of access to graphic processing unit (GPU) chips to startups. So far, four startups—Gan, Gnani, Sarvam and Soket—have been approved to build foundational AI models by Meity under the Mission. A total of over 34,000 GPUs, which are fundamental resources used to train AI algorithms on billions of parameters of data, have also been procured by Meity through cloud and data centre providers, which include the likes of Jio,Tata Communications and Yotta.
Since its announcement, the India AI Mission has been allocated close to $250 million over the previous and current fiscals. Meity does not disclose the exact budgetary utilization figure.
Leaning on language data
Singh, however, added that more than procuring GPUs and bringing down the cost of compute for startups building AI models, India's biggest efforts are being pooled into procuring public datasets on Indian languages.
'All foundational models are trained on publicly available datasets. If you look at Wikipedia, which is one of the largest sources of open data used by AI startups, there are millions of pages of data in English—but the number of pages of data in Hindi would be around a few hundred thousand. The amount of data available in other Indian languages, such as Bangla or Assamese, is even less. This is why Meity, through its Bhashini programme, conducted a datasets sourcing drive to procure language data from our own sources," he said.
Bhashini, to be sure, maintains a database of 22 Indian languages, which the AI Mission is offering to help startups build LLMs based on these languages. Singh's emphasis is already being reflected in early examples—on 8 May, Sarvam, the first startup approved by the AI Mission, introduced a text-to-speech LLM with native support for 11 Indian languages.
'In the long run, voice-first Indian language LLMs can come into application in public services—such as for a farmer who can use a public utility to reduce his irrigation costs, for access to healthcare in remote districts, and for improving education in fringe sectors. It is this that will be the biggest impact of AI in India," Singh said.
While it is too early to test most of the models, Sarvam-M, the first of the Indic language-native AI models born in India, claimed to offer 20% superior performance in vernacular languages over foreign models. However, Sarvam-M is not a foundational model—but is a 24-billion-parameter AI model trained on French startup Mistral's foundational models.
Public service apps
Industry stakeholders, too, concurred. 'For private corporates, we see a lot of investments going in with a commercial business outcome-linked approach," said Saibal Chakraborty, managing director and senior partner for India at management consultancy firm, Boston Consulting Group (BCG). 'While this only develops the upper layer of AI applications, the need of the hour is for startups to work on public service applications in AI. Not everyone needs to work on foundational models, or try to reinvent the wheel—but public utilities will likely emerge as the biggest disruptor in the long run."
The efficacy of the Mission has also been questioned recently, with large, global tech firms such as Google and OpenAI adding support for Indian languages in their latest foundational models. Kashyap Kompella, veteran AI analyst and consultant, said that there would be room for both Big Tech firms and Indian startups to coexist despite targeting a similar base.
'Enterprises are more likely to trust AI models offered by Big Tech, since they have stronger policies and safety safeguards. However, Indian firms have ample scope to disrupt critical sectors. The work on local AI models and investments should continue with a long-term focus, if the ultimate goal is not to be dependent solely on AI built outside the country," Kompella added.
Singh reaffirmed this long-term focus, adding that the Mission will seek to create an impact for the future. 'By end-2026, we hope to see the India AI Mission give birth to five foundational LLMs, and up to 100 AI applications catering to public utilities in key sectors such as agriculture, education and healthcare. The government's role is to be an enabler for private firms to take on such work," he said.