It is a speech-to-text model trained on Indian phone calls, with four transcription modes
FinTech BizNews Service
Mumbai, 28 October, 2026: Gnani AI, the Bengaluru-based frontier AI company building proprietary foundational models and enterprise agentic AI, today launched Gnani Prisma v3.0, its latest speech-to-text model. Prisma v3.0 is trained on Indian telephone audio and heavily code-mixed speech. It supports nine Indian languages and Indian English from one model and adds four transcription modes and five controls over how each transcript is written.
In Gnani AI’s evaluation using the official Vistaar scoring protocol, Prisma v3.0 recorded a macro word error rate (WER) of 16.9% across nine Indian languages, against 19.98% for ElevenLabs Scribe v2. That is about 12% lower, and Prisma v3.0 had the lower WER on each of the seven Vistaar datasets. On Vistaar audio clips longer than three seconds, automatic language identification (Auto-LID) named the spoken language correctly more than 98% of the time.
Prisma v3.0 builds on Gnani AI’s speech models, which power 50 million voice interactions every day. It is available now through Gnani’s developer API over REST, WebSocket streaming and batch.
What’s new on the Prisma stack
Prisma v3.0 gives each team the transcript it needs from the same audio. The transcription mode is set per request, so a compliance team and a voice agent can read the same call in different forms.
In Transcribe and Code-mix, inverse text normalization (ITN) converts spoken numbers into figures with Indian digit grouping. “Three lakh fifty thousand” becomes 3,50,000.
Built for how India speaks on calls
Code-mixed speech, where one sentence moves between languages, is the norm on Indian business calls. It might be Hinglish on Delhi collections call or Tanglish on a Chennai support line. These calls usually arrive over 8 kHz phone lines, and they are where transcripts most often fail the systems that depend on them. For a lender, an amount written in words is a reconciliation task. For an insurer, a misspelled product name is a gap in the audit trail.
A speaker might say "सव्वादोन कोटी" (Marathi for two and a quarter crore, or 22.5 million rupees) or "ಎರಡೂವರೆ ಕೆಜಿ" (Kannada for two and a half kilos). Speech recognition often leaves these as words, and someone has to convert them before the data reaches a CRM or a claims system. In Transcribe and Code-mix, Prisma v3.0 returns ₹2.25 crore and 2.5 kg. When a team needs the original wording, Verbatim keeps the spoken form.
“Indian business today runs on phone calls, and those calls move between languages within a single sentence,” said Ganesh Gopalan, Co-founder and CEO, 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.”
“Code-mixed speech is the default in India,” said Bharath Shankar, Chief Product and Engineering Officer, Gnani AI. “With v3.0, teams choose how each transcript is written, from a verbatim record to code-mixed text where each language keeps its own script. Spoken times like साडेदहा, Marathi for half past ten, come out as 10:30. Developers can pass up to 200 keywords per request, so their product names come back correctly.”
Availability
Prisma v3.0 is available now. Developers can try it in the Gnani APIs playground at https://app.gnani.ai by choosing Gnani Prisma v3.0 under Speech to Text. For production, it runs over REST, WebSocket streaming or batch with an API key, and is available through Gnani’s agent builder platform.
Read the technical details in the launch blog:
https://www.gnani.ai/resources/blog/prisma-v3-0-speech-recognition
