max_endpointing_delay

is a configuration parameter in LiveKit's voice agent system that sets the maximum amount of time the system will wait before deciding a user has finished speaking when it's uncertain whether they're done. • When the system's turn detector is uncertain about whether the user is done (low confidence), it waits up to

max_endpointing_delay

seconds (default: 6.0s) before triggering a response. • When the turn detector is confident, it uses the shorter

min_endpointing_delay

(default: 0.5s). • This helps avoid premature interruptions, especially for users who pause mid-sentence or in complex conversational patterns. •

max_endpointing_delay

only has effect if a turn detector model is being used. You can configure this parameter per session, for example:

session = AgentSession(
    min_endpointing_delay=0.5,  # High confidence
    max_endpointing_delay=5.0,  # Low confidence
)

The goal is to balance responsiveness with not cutting off users who haven’t actually finished speaking. Citations: #7d3a2bdc-5248-4a6f-98c3-d482a08ac091, #28b38d53-5c4a-45eb-bc3d-76f51cb57676, #a5510711-aef0-484b-8bfb-226a6ac9f504 Sources: Turn detection and interruptions | LiveKit Docs | Speech & Audio Processing | livekit/livekit_composite | DeepWiki | tests/test_agent_session.py