Hi all! I'm using deepgram's flux model for STT. Since tweaking the eot values for this model, I've noticed some strange behavior with turn detection particularly with the agent rapidly capturing user_state_changed events that switch from "listening" to "speaking". I'm not quite sure if eot is the culprit here.

def get_stt_config(self):
        return deepgram.STTv2(
            model="flux-general-en",
            eot_threshold=0.9,
            eager_eot_threshold=0.8,
            mip_opt_out=True,
        )

This is what our config looks like. Has anyone experienced similar issues? Would having the eot_threshold at max value cause issues like this?

Quickly tested with these settings using our

agent-starter-python

and I don't see the rapid switching, so perhaps there's also some other app setting influencing the behaviour.

Is using VAD with the deepgram flux model not recommended? We're currently using Silero for VAD.

I haven't used it extensively so I wanted to find you an example... we also use silero in our flux demo: https://github.com/livekit-examples/python-agents-examples/blob/main/complex-agents/note-taking-assistant/agent.py#L281C9-L281C30