Turns are driven by Deepgram's own end-of-turn (Flux/Nova) or by LiveKit's server-side VAD (Sarvam/Gemini). These control how quickly the agent decides you're done.

End-of-turn min delay: 500 ms End-of-turn max delay: 3000 ms Flux EOT threshold: default Preemptive generation (start LLM early on eager EOT)

Noise handling is separate from VAD — VAD decides when you speak; these clean the audio. Applied by your browser (WebRTC) before it reaches the agent.

Noise suppression Echo cancellation Auto gain control

For heavier, background-voice removal, LiveKit Cloud offers server-side BVC/Krisp — see the docs.