Saaras V4
Speech recognition for India's 22 languages with a 3B state-space decoder.
About
Saaras V4 is Sarvam AI's speech recognition model, announced February 11, 2026, supporting all 22 scheduled languages of India with real-time streaming transcription. It pairs an audio encoder with a 3-billion-parameter hybrid state-space language model decoder trained from scratch, a design aimed at efficiency on long audio sequences.
The model offers five output modes (transcription, translation, transliteration, verbatim, and code-mixed), keyterm prompting for brand names and jargon, speaker diarization in the batch API, and sub-150ms streaming latency for voice agents. Sarvam reports a 5.22% language-identification error rate across 22 languages on verified IndicVoices data and strong results on IndicContextEval; these are company-reported figures. It is available through Sarvam's API with Python and Node.js SDKs and LiveKit/Pipecat integrations.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 103; Trending score 78/100 from researcher assessment.