Thanks — great question. You're asking whether, when using Gemini Live alongside a separate STT model for LiveKit’s VAD, Gemini Live consumes the raw audio or the STT transcript. Short answer: • Gemini Live uses the raw audio signal. The separate STT model is used only to perform VAD/turn-detection in LiveKit; the transcript produced by that STT is not automatically fed into Gemini Live unless you implement that yourself. What that means for your setup: • Use the STT model to detect when a user has finished speaking (turn boundaries). • Once a turn is detected, the original audio stream (not the STT text) is sent to Gemini Live for processing and response generation. • If you want Gemini Live to see the STT transcript as well, you must explicitly pass that transcript into your application logic and into Gemini Live as additional input. One actionable tip: • If you want full control over turn-taking via your external STT, disable Gemini’s built-in automatic activity detection by turning off

automatic_activity_detection

in

RealtimeInputConfig

. If you’d like, tell me which STT model and LiveKit settings you’re using and I can suggest exactly where to toggle

automatic_activity_detection

and how to pass transcripts into Gemini Live if you want that behavior. Sources: Gemini Live API plugin | LiveKit Documentation