Sarvam AI's Saaras V4 Is an AI That Understands 22 Indian Languages

An Indian AI that speaks your language

Most voice AI was built for American English first and everyone else second. India's Sarvam AI is flipping that order. This week the Bengaluru company launched Saaras V4, a speech recognition model built from the ground up for how Indians actually talk — and it understands 22 Indian languages plus English.

One recording, five ways to read it

The standout trick: from a single audio clip, Saaras V4 can produce five different transcripts. A verbatim mode captures speech exactly as spoken, a transcribe mode cleans it into the native script with normalized numbers and dates, codemix keeps English words in English, translit renders Indian-language speech in English letters, and translate gives you the English meaning. The clever part is that all five come from one model — there's no chain of separate tools where each step can add its own errors.

Built for the way Indians actually speak

Saaras V4 pairs an audio encoder with a 3-billion-parameter language model that Sarvam trained in-house from scratch, and it's designed for real-world messiness: code-mixing ("yaar, meeting postpone kar do"), regional dialects, and noisy recordings. It detects the language automatically, handles dialects, and starts streaming text back in under 150 milliseconds. On Sarvam's own testing it scored best-in-class accuracy across the Vistaar benchmark for Indian languages, and the company notes that for 10 of the 22 languages, there's no commercial alternative at all.

Why it matters

This is infrastructure, not a chatbot demo. Saaras V4 ships as an API with Python and Node.js SDKs and plugs into tools like Vercel AI SDK, LiveKit Agents, and Pipecat Agents — so developers can build Hindi customer-support bots, Marathi meeting note-takers, or Tamil voice assistants that actually hear the words right. Keyterm prompting even lets you feed it names and brand terms up front so they don't get mangled. If the future of AI is voice-first — and in a country with 22 official languages, it almost has to be — models like this are how it gets built.