Thanks — good question. Yes — with the Python Agents SDK you can receive VAD-related events when someone is speaking. Two ways to get that data: • Low-level VAD metrics: Load a VAD instance (for example,

silero.VAD.load()

), attach a handler to its

metrics_collected

event, and use that VAD in your

AgentSession

. The VAD emits

vad.VADEvent

objects with speech start/stop, idle time, inference timing, etc. Example:

vad = silero.VAD.load()

def on_vad_event(event: vad.VADEvent):
    asyncio.create_task(handle_vad(event))

vad.on("metrics_collected", on_vad_event)

session = AgentSession(stt=..., llm=..., tts=..., vad=vad)

• Higher-level conversation state:

AgentSession

emits

user_state_changed

events (part of the turn-taking logic). You can listen for

ev.new_state == "speaking"

and

ev.new_state == "listening"

to know when a participant starts/stops speaking (this uses VAD internally). Example:

@session.on("user_state_changed")
def on_user_state_changed(ev: UserStateChangedEvent):
    if ev.new_state == "speaking":
        print("User started speaking")
    elif ev.new_state == "listening":
        print("User stopped speaking")

Which approach do you want to use — raw VAD metrics or the higher-level user state events? If you share a short snippet of how you initialize your agent/session, I can show the exact code to wire it into your setup. Sources: VAD Metrics | LiveKit docs | Turns overview | LiveKit docs