Sarvam introduced chatbots enabled to support voices in more than 10 Indian languages | Festina Lente - Your leading source of AI news | Turtles AI

Indian startup Sarvam AI recently launched a series of innovative technology solutions with the goal of expanding AI adoption in a complex market like India, where as many as 22 official languages and more than 19,000 dialects coexist. In such a linguistically diverse environment, Sarvam’s proposal stands out by introducing voice-enabled chatbots that support more than 10 local languages, focusing on users’ preference for voice interaction over text interaction. Sarvam AI co-founder Vivek Raghavan pointed out that writing in Indian languages is often complicated, making voice communication a more natural and accessible option for many.

Key Points
- Startup Sarvam AI introduces voice-enabled bots in more than 10 Indian languages.
- Small and open source AI models to meet local needs.
- Sarvam 2B model, based on synthetic data, reduces AI costs.
- Sarvam aligns with the Indian government’s efforts for a national AI infrastructure.

Based in Bengaluru, Sarvam AI primarily targets enterprises and businesses, offering its voice bots in areas such as customer service. A practical example is its collaboration with Sri Mandir, a startup offering religious content that has already successfully deployed Sarvam’s AI agents to handle more than 270,000 transactions. The bots developed by the startup can be integrated into different platforms, from WhatsApp to traditional voice calls, making them extremely versatile.

Backed by investors such as Peak XV and Lightspeed, Sarvam AI also announced that the cost of its voice agents will be about ₹1 ( about one cent) per minute of use, making this technology affordable even for those with limited resources. The model underlying these agents, called Sarvam 2B, is a small language model trained on synthetic data. Although the use of synthetic data is often discouraged by experts because of the risk of hallucinations and inaccuracies, Sarvam opted for this route because of the scarcity of Indian language content available online. Raghavan pointed out that the company has developed sophisticated models to clean and improve the data, thus ensuring more reliable results.

In addition to voice chatbots, Sarvam AI is launching an audio-lingual model called Shuka, built on the Saaras v1 audio decoder and Meta’s Llama3-8B Instruct model. This model will also be open source, allowing developers to use it to create advanced speech interfaces. An additional product under development is "A1," a generative AI workbench designed for lawyers, capable of consulting regulations, drafting documents and performing other legal functions.

Sarvam AI is among a small group of Indian startups that are trying to develop solutions in line with the specific needs of the local market, while contributing to the Indian government’s efforts to build its own AI infrastructure. Specifically, the government’s "IndiaAI" program aims to create a sovereign AI ecosystem that ensures data privacy and stimulates economic growth. Ongoing initiatives include the creation of a supercomputer powered by at least 10,000 GPUs and the development of language models such as Bhashini, designed to democratize access to digital services in various Indian languages.

Raghavan expressed Sarvam AI’s willingness to collaborate with the Indian government on these projects, highlighting the importance of developing technologies that meet the specific needs of the Indian population. The conclusion of this initiative is a significant step toward building an AI infrastructure that not only reflects India’s linguistic and cultural diversity, but is also accessible and functional for a wide range of users.