Overview
- Gnani.ai launched Prisma v2.5 on June 17 and made the model available to enterprise customers through APIs.
- The company says Prisma was trained on 14 million hours of proprietary Indic speech to handle dialects, background noise, compressed telephony audio, and mid-sentence code-switching without language tags.
- Gnani.ai reports internal and third-party tests showing lower word and character error rates than competitors such as ElevenLabs, Sarvam AI, Deepgram and Microsoft, though it has not published detailed benchmark methods or scores.
- The model is hosted on Indian data centres including E2E Networks to lower latency for real-time telephony and agent-assist applications, a capability Gnani.ai says makes Prisma better suited to voice-first enterprise workflows.
- The release follows Gnani.ai’s March Series B funding and is framed as part of a wider push for sovereign, locally hosted AI in India with plans to expand Prisma to markets such as Japan, the Philippines and the Middle East.