Why This AI Startup Believes Voice Bots Can Boost AI in India- Sarvam AI

Imagine your audience speaks 22 official languages and over 19,000 dialects. Does it make sense to offer an AI chatbot that works best in only a couple of languages?

Image Credits: Sarvam AI

That is something Sarvam, an Indian AI startup is working on. On Tuesday they announced a slew of new products headed including Voice Enabling AI Bot capable over 10 Indian regional languages. He does not think that text-based chat is necessary, asserting: “In India our preference to talk an AI in a local language.

Sarvam is also introducing a mini LM, an autezhöhlr than lawyers AI tool and LMa (audio language model).

“Everybody speaks their own language. Vivek Raghavan, the co-founder of Sarvam AI says that “Typing in Indian languages is really hard today,” told TechCrunch.

A Bengaluru-headquartered startup, Yellow Messenger provides AI voice-based bots to businesses across customer support and other industry. One such payment segment is of a content startup named Sri Mandir, which creates religious videos and other spiritual material has leveraged AI agent for payments and made around 270k transactions.

AI voice agents: With Sarvam, AI Voice agents can be used from WhatsApp to app and even on traditional calls.

Image Credits: Sarvam AI

Supported by Peak XV and Lightspeed, the company intends to charge as little ₹1 ($0.013) per minute of use for its AI agents.

To develop the voice-operated AI agents, this startup used an artificial general intelligence model named ascore for which they created a deep learning framework called Sarvam 2B with small language constructs that was trained on 4 trillion tokens dataset. The model here uses synthetic data, i.e., the data created by a large language model to look like real-world training examples. But AI specialists caution that there may be a downside to this: models are liable to fabricate false facts while the use of artificial data. To tackle this, Sarvam began working on either cleaning the synthetic data or processing it further to make the data better.

Raghavan claimed that Sarvam 2B will be much cheaper than similar models in the industry and that the startup is open-sourcing the model to allow the community to build on it.

“Large language models are exciting, but smaller models can offer better experiences, lower costs, and less delay,” Raghavan explained. He added that while large models are suitable for occasional queries, smaller models are better for situations requiring millions of daily interactions.

Sarvam is also releasing an audio-language model called Shuka, built using its Saaras v1 audio decoder and Meta’s Llama3-8B Instruct. This model will also be open-sourced so developers can create voice interfaces using the startup’s translation, text-to-speech, and other modules.

Another product, “A1,” is a generative AI tool designed for lawyers. It can look up regulations, draft documents, redact them, and extract data.

Sarvam is one of the few Indian startups promoting AI that aligns with the country’s needs and supports the government's efforts to build its own AI infrastructure.

Governments worldwide are increasingly focusing on "sovereign AI"—AI developed and controlled at the national level to protect data privacy, boost economic growth, and tailor AI to cultural contexts. The U.S. and China are leading in this area, and India is following with its “IndiaAI” program and language-specific models.

One initiative under the IndiaAI program, called IndiaAI Compute Capacity, aims to create a supercomputer powered by at least 10,000 GPUs. Another project, Bhashini, seeks to make digital services accessible in various Indian languages.

Raghavan said Sarvam is ready to support the IndiaAI program. “If the opportunity arises, we will work with the government,” he said in the interview.