For hundreds of millions of Indians, the internet still speaks the wrong language. English-first apps, text-heavy screens and voice assistants trained on faraway accents leave out the very people who need digital services most — the first-time smartphone user in a village in Bihar, the domestic worker sending money home, the grandmother trying to watch her favourite show on a smart TV. Navana.ai, a Mumbai startup founded by brothers Raoul and Jai Nanavati, has spent eight years building the technology that closes that gap: Bodhi, a voice AI stack trained from the ground up on how India actually speaks.
Built for how India actually speaks
Bodhi is Navana’s family of proprietary speech models, and it is built for Indian realities rather than adapted to them. The company says the models cover 12 Indian languages and 45 dialects, trained on hundreds of thousands of real-world conversations. They are designed to handle the things that trip up global voice systems in India: background noise, overlapping speakers, code-switching between languages mid-sentence, and the sheer variety of regional accents. The company says Bodhi is around five times smaller than the nearest comparable architecture, which lets it run on ordinary CPUs as well as GPUs — and lets a bank run the whole stack inside its own data centre.
Bodhi TTS and Bodhi STT
In September 2026 Navana opened Bodhi TTS, its text-to-speech model, to the public. It speaks 10 Indian languages — Hindi, Telugu, Tamil, Marathi, Bengali, Odia, Malayalam, Kannada, English and Gujarati — in more than 50 voices, and the company says it starts generating audio in under 100 milliseconds. It is priced at ₹12 per 10,000 characters, roughly ₹0.72 a minute of speech, and the company says that works out to about one-eighth of ElevenLabs’ price and 60% below Sarvam’s Bulbul v3. A developer can clone a voice from a few seconds of reference audio, and teams can correct how names and numbers are pronounced without retraining the model. The model is tuned for the text Indian businesses actually speak: currency amounts, phone numbers, PAN numbers, dates, loan instalments and scheme names. Alongside it, Bodhi STT — speech recognition across 10 languages and more than 40 dialects — opened by invitation, with its benchmark results published on Hugging Face.
From a grandmother’s TV to a hundred million minutes
The idea behind Navana started at home. The brothers watched their grandmother struggle to navigate a smart TV, domestic workers wrestle with banking apps while sending money home, and cab drivers ask for help downloading apps — and travelled across India to understand the problem first-hand. The hypothesis they came back with was simple: Indians prefer voice over text, especially in their own language. The company spent a year doing field research with its first design partner, Ujjivan Small Finance Bank, interviewing around 500 customers across 10 states and 15 languages. The result is visible in the bank’s “Hello Ujjivan” app, which the company says now serves more than 1.5 million users in 12 Indian languages. Navana also worked with IISc Bengaluru on RESPIN, open-sourcing a 10,000-hour Indian-language speech corpus spanning 45 dialects — one of the largest openly available Indic speech datasets anywhere.
Today the company says its systems have processed more than 100 million voice AI minutes for customers including Bajaj Finserv, Protean, Ujjivan Small Finance Bank, Jana Small Finance Bank and Grihum Housing Finance, across sales, collections, compliance and support — and that more than 40 enterprises now run on its platform.
Sovereign by architecture
Navana describes Bodhi as sovereign by architecture, not merely by hosting: the company trains and owns the entire speech stack end to end, with no customer audio routed through a third-party model provider. The platform is built for regulated industries, with on-premise deployment, ISO 27001 and SOC 2 Type II certification, and alignment with RBI guidelines and data-localisation norms. The work has drawn collaborations with IISc Bengaluru, IIT Madras, Microsoft Research and the Gates Foundation. In September 2026 the company raised ₹40 crore in a Series A round led by entrepreneur and investor Ronnie Screwvala, to deepen its presence across banking, financial services and insurance and to build out its AI Contact Centre and automated evaluation platform.
More Read
The next billion Indians will not learn to type in English to use the internet — the internet will learn to listen in their language. Navana is building the ears for that future, one conversation at a time, in twelve languages and forty-five dialects.
Keep reading: Ask the Manuscripts: Inside BharatiyaGPT, the Nagpur AI Turning India’s Ancient Texts Into Conversation · India’s First National-Security Fab: Inside Bharat Semi, the Delhi Startup Building the Sensors India Couldn’t Buy · 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.
