India does not talk in neat, single-language sentences. A customer switches from Hindi to English halfway through a complaint, drops in a brand name, speeds up, and does it all over a crackly phone line. Most speech recognition still stumbles on exactly that moment. FonadaLabs, the speech AI division of Noida-based Fonada, has built a model for the way India actually talks — and the company says it now leads the world at understanding Hindi.
Kapture V3 is FonadaLabs’ speech recognition model, live on its API since 23 October 2026. It turns conversational audio — call-centre recordings, voice notes, voice-agent calls — into verbatim transcripts in the speaker’s own script, with segment-level timestamps. It handles ten Indian languages alongside English: Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu.
Built for the way India actually talks
The design brief behind Kapture V3 starts from a simple observation: Indian speech is unscripted, fast, and constantly mixing languages mid-sentence, often over a phone line. A model trained mostly on clean, read-out English audio will miss the words that matter most in a real conversation. FonadaLabs built Kapture V3 for the opposite conditions — noisy, natural, code-mixed speech — and tuned it for the telephony and voice-agent workloads where Indian businesses actually need it.
That focus shows up in the practical details. There is no limit on audio length: long recordings are split and transcribed in parallel, so an hour-long call does not have to wait for an hour of processing. The company says the median server processing time for a short clip is about 0.1 seconds. And because transcripts arrive in each language’s native script, with timestamps marking every segment, they can drop straight into a contact-centre workflow, a search index, or a voice agent’s memory.
No. 1 for Hindi, on the world’s open leaderboard
On 4 November 2026, FonadaLabs announced that Kapture V3 has secured the No. 1 position for Hindi on the Hugging Face Open ASR Leaderboard, the open, community-driven benchmark where speech models from around the world are compared on the same tests. The company says the model recorded an average word error rate of 6.54% on the Hindi benchmark, placing it ahead of the other leading global models on the board — the same arena where OpenAI, Google, Microsoft, NVIDIA and IBM all compete.
It is a striking result for a model built specifically for Indian languages. Hindi is among the world’s most spoken languages, yet global speech models have historically treated it as an afterthought, trained on far less data and tested far less rigorously than English. Kapture V3’s own published results point the same way: on the Voice Arena Monsoon Hindi test, the company reports an observed word error rate of 6.21%, and on the ARTPARK-IISc Vaani Benchmark for Hindi, 13.1%. Across a 10-language sample of the Vistaar benchmark, it reports 17.43%.
More Read
Eleven languages, one API
Kapture V3 arrives as an API, not a download: developers send audio and get back text, in eleven languages, through a single endpoint. That simplicity is deliberate. Fonada’s wider platform is voice infrastructure for Indian telecom networks — telephony hosting, speech recognition, text-to-speech and noise cancellation, stitched into one pipeline so teams can build conversational agents without managing each layer themselves. Everything is hosted in India, a point the company stresses for enterprises that need their voice data to stay within the country.
The applications reach well beyond the call centre. Accurate, affordable transcription in Indian languages unlocks searchable customer conversations, voice-enabled services for people who prefer speaking to typing, and voice agents that can finally keep up when a caller switches languages mid-sentence.
Priced for real-world use
Alongside accuracy, FonadaLabs has pushed hard on cost. The company prices Kapture V3 inference at approximately ₹0.10 per minute of audio — about ₹6 per hour — positioning it as a practical option for businesses that want to transcribe at scale rather than ration every minute. In a market where global providers charge premium rates for speech services, that pricing is aimed squarely at lowering the barrier for Indian-language adoption.
The company also acknowledged the support of the Ministry of Electronics and Information Technology and the IndiaAI Mission, saying a 40% compute subsidy supported its model training and experimentation — a reminder of how public compute infrastructure is now feeding directly into Indian model-building.
What comes next
Kapture V3 is one part of a wider roadmap. FonadaLabs says it will continue research and development in natural speech generation, Indian-language pronunciation and efficient inference, with text-to-speech, voice cloning, noise cancellation and AI dubbing all on the way. The goal, in the company’s words, is speech technology that understands Indian languages accurately while remaining practical and affordable for real-world enterprise use.
“Reaching the top position on the Hugging Face Open ASR Leaderboard is an important milestone for FonadaLabs and, more importantly, for the Indian speech AI ecosystem,” said Nitish Gopalani, Co-Founder and CEO of Fonada. “Our focus has always been to build speech technology that understands Indian languages accurately while remaining practical and affordable for real-world enterprise use.”
A model that hears Hindi best, understands a caller who mixes languages without missing a beat, and does it for a few rupees an hour is not just a benchmark win. It is the kind of quiet infrastructure that lets millions more Indians talk to technology in the language they actually speak — and be understood, word for word.
Keep reading: Decisions in 44 Milliseconds: Inside Sarvam’s Prism, India’s Open-Weight Decision Model · Pay Only for the Tokens You Use: Inside E2E Networks’ Blaze, India’s Pay-Per-Token AI Inference Service · 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.
