India does most of its business on the phone, and it does it in dozens of languages, accents and dialects. A customer in Indore may switch mid-sentence from Hindi to English, a caller in Coimbatore may speak fast Tamil over a crackly line, and a sales call in Jaipur may run for twenty minutes before anyone talks about price. Bengaluru-based Vaani AI Research is building the voice technology for exactly that world — its own speech models, its own inference stack, and a self-serve platform that lets any team put a voice agent on a real phone line.
On 29 October 2026, the company opened that platform to everyone. Developers, startups and product teams can now describe the agent they want, preview its voice, configure what it can do, and deploy it across phone calls, websites, apps and WhatsApp — without assembling the speech, telephony and orchestration layers themselves.
The problem with stitched-together voice agents
A production voice call today usually runs on four separate layers that were never designed to work together: telephony, speech-to-text, a reasoning layer, and text-to-speech. Each layer comes from a different vendor, and the joins between them are where conversations go wrong — metallic voices, dead air, awkward pauses, and agents that lose the thread when a caller interrupts or changes tack.
Vaani’s answer is to own the whole chain. The company says it builds its own data-efficient speech models alongside a specialised inference stack and an orchestration platform, so the pieces are designed to work as one system rather than bolted together. The result, the company says, is a voice agent that sounds natural, holds context across a long conversation, and keeps talking like a person rather than reading like a script.
One harness, from model to phone line
At the centre of the platform is what Vaani calls its voice harness — the infrastructure layer that connects its speech models with everything a live call needs. It brings together inference optimisation, memory and context management, conversational orchestration, telephony, tool integrations and human-in-the-loop workflows, so a team does not have to stitch those pieces together itself.
That harness is also what makes the platform genuinely self-serve. Teams can build an agent through APIs, or use the configurable harness to design the call flow in plain language: how the agent handles objections, interruptions and edge cases; what actions it can take, from holding a call to scheduling a callback or handing over to a human; and which knowledge base it should answer from mid-call. Once it is live, every call ends with clean, structured data — extracted fields and a disposition — so nobody has to read through transcripts to know what happened.
More Read
The company says the platform is built for speed as well as control. Its multilingual engine runs at a p90 latency of around 450 milliseconds, and its agents are designed for zero dead air and natural turn-taking — pausing when it matters, making small talk while a tool runs, and switching language or dialect automatically as the caller does.
Built for the languages India actually speaks
Vaani’s founding idea is that voice AI should start from the languages where data is scarce, not treat them as an afterthought. The company says its models are capable of speaking in 22 or more Indian languages, and the platform is designed for the messy reality of Indian conversation: accents the training data barely covered, code-switching between languages, and long consultative calls in sales, financial services and healthcare where nuance matters.
That focus runs through the whole product. Agents can auto-switch to the caller’s language and dialect, carry realistic voices with emotional tags, and keep the context of a conversation alive across calls — remembering, for example, that a customer prefers Hindi and wants updates on WhatsApp rather than SMS. It is a small detail, and it is exactly the kind of detail that decides whether a caller feels heard or handled.
Teacher AI: describe it, and watch it build
The newest part of the platform is Teacher AI, Vaani’s guided way of building an agent. A team describes the agent in its own words — who to call, what to say, and what counts as success — and Teacher AI asks clarifying questions about languages, objections, qualifying criteria and handoffs. Each answer tightens the agent before it ever makes a call.
Then, the company says, it builds in seconds: the prompt, voice, language mix and tools come together live, with a readiness score as it goes. If a later tweak changes the prompt, the team can review the difference before anything is saved, and keep refining by chat — changing the tone, adding a language, or teaching the agent a new objection — without starting over. The result is a deployment-ready agent, not a draft to finish.
The team behind the voice
Vaani was founded by Tushar Shinde, Nitesh Tripathi and Nitish Mishra, with roots in speech research at the Indian Institute of Science, Bengaluru, and the Indian Institute of Technology Madras. Shinde worked on reinforcement learning at IISc before building enterprise language systems and products; Tripathi worked on voice and natural language processing at IISc; and Mishra came from IIT Madras. The company says its self-serve platform grew out of its work deploying voice AI inside live enterprise workflows, where every call shows what breaks — and that the harness was hardened there before being opened to everyone.
The company is backed by Venture Catalysts and the Google AI Accelerator, alongside an undisclosed institutional investor, and its launch topped Product Hunt’s Product of the Day, the company says — an early sign of how eagerly developers have been waiting for a voice stack that treats Indian languages as the default, not the edge case.
From a single phone call to a thousand, Vaani is betting that the next great Indian software story will not be read on a screen at all. It will be heard — in the caller’s own language, at human speed, with a voice that knows when to pause, when to help, and when to hand over. That is a future worth listening to.
Keep reading: Meaning First, Not Word by Word: Inside NeuroDX’s Sphota, India’s Brain-Inspired AI Model · The Carbon You Can’t See: Inside Fitsol, the Gurugram Startup Turning India’s Supply Chains Green · More Industrial Automation & Technology stories
Like stories like this? Add Exceptional Technology as a preferred source on Google to see more of our reporting in your search results and Discover feed.
