Gnani.ai launches Prisma v3.0 speech model
By Abha Lohia ·
Gnani.ai: funding and every story →
- Gnani.ai is a Bengaluru AI company building its own foundation models and enterprise voice agents, handling about 50 million voice interactions daily
- Its new Prisma v3.0 converts speech to text, trained on Indian phone calls with mixed languages, across nine Indian languages plus Indian English
- The company says it beat a rival model on error rate, 16.9% against 19.98%, and offers it through a developer API
What happened
- Prisma v3.0 is a speech-to-text model for Indian call audio.
- It covers nine Indian languages and Indian English in one model.
- It has four transcription modes and five controls over transcript style.
- Its word error rate (share of words transcribed wrongly) is 16.9% across languages, by the company's own test.
- It is available through REST, WebSocket streaming and batch API.
Why it matters
- Indian phone calls mix languages, which general models handle poorly.
- Different teams, such as compliance and voice agents, can read one call in different forms.
- The error-rate claim is the company's own evaluation.
The business
- Gnani.ai builds proprietary foundation models.
- It sells enterprise agentic AI and voice agents.
- Its speech models power around 50 million voice interactions a day.
The founders
In numbers
16.9%Word error rate · 2026
Up against
ElevenLabs
Everything on Gnani.ai → · All startup news on 28 October 2026 →
Is this Gnani.ai? Get your “Featured on StopDown” badge →