Launch · AI & Deep Tech

Sarvam AI releases Saaras V4 speech model

By Startup Enthusiast · 25 September 2026

What they do

Sarvam AI is an Indian full-stack sovereign AI platform

It released Saaras V4, a speech recognition model supporting 22 Indian languages and English

Saaras V4 can produce five output modes from the same audio without extra pre-processing

Sarvam AI released Saaras V4, a speech recognition model.
Saaras V4 supports 22 Indian languages and English.
It can produce five output modes from the same audio: verbatim, normalised, code-mixed, transliterated, and translated.
All five formats come from one model, avoiding cascading errors from extra pre-processing steps.
Saaras V4 combines an audio encoder with a 3-billion-parameter hybrid state-space language model trained from scratch in-house.
It is designed to handle code-mixing, dialect variations, and noisy audio.
Saaras V4 supports real-time speech streaming with time to first token below 150 ms and can process multi-minute recordings in under a second.

Saaras V4 recorded a 5.22% language-identification error rate across all 22 Indian languages using verified IndicVoices data.
It achieved a 2.9% language-identification error rate across the 10 most widely spoken Indian languages.
Saaras V4 achieved SOTA performance on all 22 Indian languages in its evaluations.
It achieved the lowest average word error rate across seven English speech-recognition benchmarks.

Sarvam AI offers Voice Agents, Text to Speech, Speech to Text, Document Digitisation, and Translation APIs.
Text to Speech supports natural voices across 11 Indic languages.
Speech to Text supports high-accuracy ASR across 12 Indic languages.
Translation supports state-of-the-art translation across 23 languages.
Sarvam AI has served over 10 billion tokens with median latency under 100 ms and 99.9% uptime SLA.
It offers deployment options: Sarvam Cloud, Private Cloud (VPC), and On-Premises.
Sarvam AI is SOC 2 Type II, ISO 27001, and DPDP compliant with role-based access, audit trails, and data residency controls.

Product: Foundation models and agentic AI systems with voice capabilities and enterprise applications
Customers: Enterprise customers across finance, aviation, telecom and government sectors
How it makes money: SaaS and licensing model for AI models and enterprise AI applications
What sets it apart: India-built sovereign AI stack with focus on cost-effective deployment and localization