You asked whether you can use Deepgram Flux as the turn detector while using Nova-3 as the main STT. Short answer: no — LiveKit Agents does not let you split turn detection and transcription across two different STT models. Why: • Turn detection in

stt

mode uses endpointing signals from the same STT stream (the model you configured for transcription), so you cannot point endpointing to Flux while Nova-3 handles transcripts. • Turn detection alternatives are:

vad

(silence/VAD),

realtime_llm

, or a dedicated/custom turn detector plugin (e.g., LiveKit’s turn detector). What you can do instead (options): • Use Nova-3 for STT and enable a separate turn detector plugin (MultilingualModel or a custom turn detector) for better turn-taking. • Use VAD-based detection alongside Nova-3 for simpler endpointing behavior. If you want, tell me which option sounds best (Nova-3 + plugin or Nova-3 + VAD) and I’ll provide example configuration code. Sources: Agents (Python) | livekit/livekit_composite | DeepWiki | examples/voice_agents | livekit/livekit_composite | DeepWiki | LiveKit turn detector plugin | LiveKit docs | Turn detection and interruptions | LiveKit Docs