Sarvam AI closed a $50 million Series B led by Lightspeed and Peak XV to build India’s first voice-native large language model. The Bangalore-based startup is betting that text-first AI misses 400 million Indians who prefer voice interaction.
The Funding Details
Round size: $50 million Series B
Lead investors: Lightspeed India, Peak XV Partners
Valuation: $300 million (post-money)
Total raised: $72 million
What They’re Building
Sarvam’s core product is a voice-native LLM trained on 22 Indian languages. Unlike competitors who bolt speech-to-text onto existing models, Sarvam processes audio directly—eliminating the latency and errors from transcription.
Key technical claims:
- 300ms response latency (vs. 800ms+ for STT + LLM pipelines)
- 94% accuracy on Indian accents (vs. ~78% for global models)
- Support for code-switching (Hindi-English, Tamil-English, etc.)
The Market Opportunity
India has 400 million+ users who are voice-first—primarily due to literacy barriers, language preferences, or hands-free contexts. This includes:
- Rural users (68% prefer voice over text)
- Blue-collar workers (limited typing proficiency)
- Drivers, delivery workers, field staff (hands occupied)
Competition Landscape
Sarvam competes with Bhashini (government-backed), Reverie Language Technologies, and international players adapting for India. Their edge: purpose-built architecture vs. adapted models.
For companies building voice AI agents for the Indian market, Sarvam’s APIs could provide the foundational layer for localized deployments.
What to Watch
Sarvam plans to launch an enterprise API in Q2 2026. Early partners include a major telecom, two banks, and a government department. If the accuracy claims hold in production, this could reshape how Indian businesses deploy conversational AI.