VAD and turn detection configuration guide

LiveKit Agents provides extensive control over Voice Activity Detection (VAD) and turn detection. However, not all parameters apply to every agent configuration. This guide explains which parameters matter for your specific setup.

VAD Configuration Parameters

These parameters are defined in the Silero VAD implementation and control how speech is detected in audio streams:

These parameters are used whenever VAD is present in your pipeline.

Turn Detection and Interrupt Parameters

These parameters are defined in the AgentSession constructor and control how the agent handles conversational turns:

Scenario Reference

Pipeline Agents with Turn Detector Model

When using a turn detector model alongside VAD and STT, all parameters are active:

The turn detector model provides confidence scores that influence timing between min_endpointing_delay and max_endpointing_delay.

Pipeline Agents without Turn Detector Model

Without a turn detector model, max_endpointing_delay becomes irrelevant since there are no confidence predictions to act on:

Realtime Model Agents (OpenAI, Google)

When using realtime models like OpenAI's Realtime API or Google's realtime models, VAD and turn detection are handled server-side:

Universal Parameters

These parameters work across all agent types:

Quick Reference by Agent Type

How Turn Detection Mode is Selected

The turn detection mode is automatically selected based on available components, with this priority order:

1. realtime_llm — Uses server-side VAD and turn detection
2. vad — Uses client-side Silero VAD
3. stt — Falls back to STT-based detection
4. manual — No automatic turn detection

Configuration Examples

Pipeline Agent with Turn Detector

from livekit.agents import AgentSession
from livekit.plugins import silero

session = AgentSession(
    vad=silero.VAD.load(
        min_speech_duration=0.1,
        min_silence_duration=0.3,
        prefix_padding_duration=0.5,
        activation_threshold=0.5,
    ),
    turn_detector=my_turn_detector,
    allow_interruptions=True,
    min_interruption_duration=0.5,
    min_endpointing_delay=0.5,
    max_endpointing_delay=6.0,  # Used with turn detector confidence
)

Pipeline Agent without Turn Detector

from livekit.agents import AgentSession
from livekit.plugins import silero

session = AgentSession(
    vad=silero.VAD.load(
        min_speech_duration=0.1,
        min_silence_duration=0.5,  # More important without turn detector
        prefix_padding_duration=0.5,
        activation_threshold=0.5,
    ),
    allow_interruptions=True,
    min_interruption_duration=0.5,
    min_endpointing_delay=0.8,  # Acts as the fixed delay
    # max_endpointing_delay not needed
)

Realtime Model Agent

from livekit.agents import AgentSession
from livekit.plugins import openai

session = AgentSession(
    llm=openai.realtime.RealtimeModel(),
    # VAD params not needed - handled server-side
    # Endpointing delays not needed - handled server-side
    allow_interruptions=True,
    min_interruption_duration=0.5,
)

Summary

Understanding which parameters apply to your agent configuration prevents confusion and helps you tune the right settings. Remember:

• Pipeline agents give you full control over VAD and turn detection
• Realtime agents delegate most detection to the server, but you still control interruption behavior
• Turn detector models enable confidence-based timing with the endpointing delay range

For complete documentation on turn detection modes, interruption handling, and session configuration, see the Turns overview in the LiveKit docs.