If your target market has 22 official languages and its people speak over 19,000 dialects, does it make sense to offer a text-only AI chatbot that works best in some languages?
That's the problem Indian AI startup Sarvam has been trying to solve, and on Tuesday it unveiled a suite of products including a voice-enabled AI bot that supports more than 10 Indian languages, as it believes people in India would prefer to talk to an AI model in their own language rather than chatting over text. The company also unveiled a compact language model, an AI tool for lawyers, and a spoken language model.
“People prefer to speak in their own language. Today, it is very difficult to type in Indian languages,” Sarvam AI co-founder Vivek Raghavan told TechCrunch.
The Bengaluru-based startup primarily targets businesses and corporations, pitching its AI voice-enabled bots to various industries, especially those that require customer support. As an example, the company cited one of its clients, religious content startup Sri Mandir, which uses Sarvam's AI agents to accept payments and has processed over 270,000 transactions so far.
The company says its AI voice agent can be deployed within the WhatsApp app and will also work over traditional voice calls.
Backed by Peak XV and Lightspeed, Sarvam plans to charge a starting price of 1 rupee (about 1 cent) per minute for using its AI agents.
Image credit: Sarvam
The startup is building its voice-enabled AI agent on top of a foundational small-scale language model called Sarvam 2B, which was trained on a dataset of 4 trillion tokens, and Raghavan said the model was trained entirely on synthetic data.
AI experts often advise caution when using synthetic data (essentially data generated by large language models that aim to replicate real-world data) to train other AI models, as LLMs tend to hallucinate and fabricate information that may not be accurate, and training an AI model on such data can exacerbate the inaccuracies.
Raghavan said Sarvam chose to use synthetic data because there is very limited Indian language content available on the open web, adding that the startup developed models to clean and improve the data it initially used to generate the synthetic dataset.
The founders claim that Sarvam 2B will cost one-tenth of its industry counterparts, and the startup is open-sourcing the model in the hope that the community can develop it further.
“Large language foundation models are very attractive, but smaller language models can provide a better, more specific, less costly, and less-latent experience,” says Raghavan. “If you run one or two queries a week or a month, you should use a large language model. But for use cases that require millions of interactions every day, I think smaller models are a better fit.”
The startup has also released an audio language model called Shuka, built on the Saaras v1 audio decoder and Meta's Llama3-8B Instruct, which has also been open-sourced so developers can build voice interfaces using the startup's translation, TTS, and other modules.
Additionally, they have another product called “A1,” which is a generative AI workbench designed for lawyers that can look up regulations, draft documents, edit, extract data, and more.
Sarvam is among a small group of Indian startups that are putting forward use cases that align with the country's interests and contributing to the government's efforts to develop indigenous AI infrastructure.
Governments around the world are increasingly pursuing “sovereign AI,” AI infrastructure developed and controlled at the national level. These efforts aim to protect data privacy, promote economic growth, and align AI development with cultural context. Currently, the United States and China are investing the most in this area, followed by India with its “IndiaAI” program and language-specific models.
One of the initiatives under the IndiaAI program is called IndiaAI Compute Capacity, which plans to build a supercomputer with at least 10,000 GPUs. One of the models being developed is called Bhashini, which aims to democratize access to digital services in various Indian languages.
Raghavan said his startup is ready to contribute to the IndiaAI program. “If the opportunity arises, we will work with the government,” he said in an interview.