Audio / Speech-to-Text

Saaras V4

Speech-to-text model from Sarvam AI · released 2026-08-18 · vendor-reported (not yet ranked on an independent leaderboard)

Unrankednot on an independent leaderboard yet
2026-08-18release

Specifications

Languages
22 Indian languages + English
Architecture
Audio encoder + 3B-parameter hybrid state-space language model (per launch coverage)
Output modes
transcribe, translate, verbatim, translit, codemix
API model id
saaras:v4

Notes

  • Not ranked on the Artificial Analysis speech-to-text leaderboard at the time of writing; accuracy claims are Sarvam's own.

How this is measured

AA-WER v2 (Speech-to-Text)

Word error rate, % of words transcribed incorrectly (lower is better)

Artificial Analysis Word Error Rate index for non-streaming transcription. AA-WER v2 combines three datasets: AA-AgentTalk (50%), VoxPopuli-Cleaned-AA (25%) and Earnings22-Cleaned-AA (25%). Measured independently on dedicated hardware.

Release & sources

Release
2026-08-18
Release source
Sarvam changelog, 18 Aug 2026 (a press article dated 23 Sep 2026 also covered the launch)

More STT models from Sarvam AI