mysterious-noon-79219
11/21/2025, 9:38 AMmysterious-noon-79219
11/21/2025, 10:14 AMinput_speech_started) is being emitted when the realtime's model-sided VAD detects an interrupt but it's not clear whether this is the only way this event is used because the name is a bit confusing and it sounds like the event is overloaded.
For example, here I can see it being used when Gemini live api detects an interruptlively-winter-88668
11/21/2025, 1:21 PMinput_speech_started from my agent but as you saw, it fires for BOTH interrupts and user starting to speak:
# Access the RealtimeSession through the session's internal activity
# The input_speech_started event is from server-side VAD in the realtime model
# Note: This event may only fire when server-side VAD is active
if hasattr(session, '_activity') and session._activity:
rt_session = getattr(session._activity, '_rt_session', None)
if rt_session:
@rt_session.on("input_speech_started")
def on_input_speech_started():
print("**** Input speech started")
else:
print("Warning: Could not access RealtimeSession. Event may not fire if using client-side VAD.")
else:
print("Warning: Could not access session activity.")