Listen to any business call in India and you will hear something no speech-recognition model built abroad was designed for: a sentence that starts in Hindi, borrows an English word halfway through, and lands a number in Marathi. India does not speak one language at a time — it mixes them, mid-sentence, on noisy phone lines, every single day. A Bengaluru company has spent a decade teaching machines to hear exactly that, and its newest model is its sharpest listener yet.
Gnani AI, founded in 2016 by Ganesh Gopalan and Ananth Nagaraj and based in Bengaluru, launched Prisma v3.0 on 28 October 2026. It is the company’s latest speech-to-text model, built for Indian telephone audio and heavily code-mixed speech, and it turns the calls Indian businesses actually run on into transcripts they can actually use.
Gnani AI’s brand artwork (official company imagery).
Built for the way India actually talks
Most speech models are trained on clean studio recordings. Prisma v3.0 is trained on Indian telephone audio — the real thing, with background noise, regional accents and speakers switching languages inside a single sentence. It takes native 8 kHz telephony audio as well as 16 kHz audio from apps and internet calls, and it handles ten languages from a single model: Hindi, Bengali, Marathi, Gujarati, Punjabi, Kannada, Telugu, Tamil and Malayalam, plus Indian English.
That single-model design matters because code-mixed speech is the default in India, as the company’s chief product and engineering officer Bharath Shankar puts it. Before Prisma can write a transcript, its automatic language-identification layer has to work out which language is being spoken — and the company says it names the spoken language correctly more than 98 per cent of the time on audio clips longer than three seconds.
More Read
One call, four transcripts
Prisma v3.0’s most elegant idea is that a correct transcript is not always a usable one. A compliance team, a voice agent and a data scientist all need different versions of the same call, so the model offers four output modes, chosen per request. Verbatim keeps every filler word and repetition, in native script, for audit and compliance records. Transcribe produces clean, readable text with amounts, dates and percentages turned into digits. Code-mix writes each word in the script of the language it was spoken in, keeping English terms in English. Transliterate renders the whole transcript in English letters, whatever language was spoken.
Take a missed-EMI call: “हाँ सर, वो… मेरा ईएमआई पाँच तारीख को काटना था, बट अभी तक पेमेंट रिफ्लेक्ट नहीं हुआ।” Verbatim keeps it word for word; Transcribe turns it into clean text with the figure as ₹12,500; Code-mix keeps the Hindi, Kannada and English each in its own script. The words stay in the speaker’s language — only the writing changes. Developers can also pass up to 200 keywords per request, so product names and technical terms come back spelled correctly.
Numbers that sing
In the company’s evaluation using the official Vistaar scoring protocol, Prisma v3.0 recorded a macro word error rate of 16.93 per cent across nine Indian languages, against 19.98 per cent for ElevenLabs Scribe v2, 20.51 per cent for Microsoft Azure Speech-to-Text and 26.57 per cent for Deepgram Nova 3. The company says Prisma had the lower error rate on each of the seven Vistaar datasets, and a lower rate than both ElevenLabs and Azure in all nine Indian languages. It also reports that false text generated from noise-only audio fell from 13.4 per cent with the previous model to 4.2 per cent — so silence stays silent.
“Indian business today runs on phone calls, and those calls move between languages within a single sentence,” says Ganesh Gopalan, co-founder and CEO of Gnani AI. “Prisma v3.0 gives enterprises the ability to transcribe them the way they were spoken, with the numbers and names a business needs to act on.”
From a decade of Indian voice
Prisma v3.0 builds on speech models the company says already power 50 million voice interactions every day, serving more than 200 enterprises. Gnani AI raised a $10 million Series B round in March 2026 led by Aavishkaar Capital with Info Edge Ventures, and it has been building sovereign voice technology alongside Prisma — including Inya VoiceOS, its voice-to-voice foundation model, and the Vachana speech models. Prisma v3.0 is available now through the company’s developer API over REST, WebSocket streaming and batch processing, so a team can try it on a single call today and run it across a whole contact centre tomorrow.
That is the quiet triumph here: a Bengaluru team looked at the most ordinary, gloriously messy thing about Indian conversation — the way we blend languages without thinking — and built a machine that finally hears it the way we speak it. Every call transcribed faithfully, in every Indian language, is one more corner of the country’s business that works in its own voice. Prisma v3.0 is Indian speech technology, built for India first — and it is only just getting started.
Keep reading: The Mini AI Server Inside Your Phone: Inside NimbleEdge, the Bengaluru Startup Putting AI on the Device · Before the Water Rises: Inside Aurassure, the Bhubaneswar Startup Building India’s Climate Early-Warning Network · 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.
