Sarvam's Saaras V4 transcribes 22 Indian languages

TL;DR

• Sarvam published Saaras V4 on 24 August: a speech-to-text model for all 22 Indian languages and English, including Indian-accented English.
• One model gives five outputs (verbatim, transcribe, codemix, translit, translate), takes up to 50 key terms, and streams with under 150 ms to the first token, Sarvam says.
• It's the default model on Sarvam's API, with REST, batch and WebSocket access and SDKs for Python and Node.