To configure end-of-utterance (EOU) values in LiveKit, you can adjust several key parameters that control when the system determines a user has finished speaking. Below are the main configuration options with code examples: Basic EOU Configuration:

from livekit.agents.voice import AgentSession

session = AgentSession(
    turn_detection="vad",  # or custom model
    min_endpointing_delay=0.5,  # High confidence delay (default: 0.5s)
    max_endpointing_delay=6.0   # Low confidence delay (default: 6.0s)
)

Turn Detection Modes: •

"vad"

- Uses VAD silence detection (default fallback) •

"stt"

- Uses speech-to-text end-of-speech signals •

"realtime_llm"

- Uses server-side detection from realtime models • Custom model - Uses specialized EOU detection models Custom Turn Detector Implementation:

from livekit.plugins.turn_detector.multilingual import MultilingualModel
from livekit.agents import llm

class CustomTurnDetector:
    def __init__(self, userdata=None):
        self._model = MultilingualModel()
        self.last_eou_probability = None
        self.userdata = userdata

    async def predict_end_of_turn(self, chat_ctx: llm.ChatContext) -> float:
        probability = await self._model.predict_end_of_turn(chat_ctx)
        self.last_eou_probability = probability
        print(f"EOU Probability: {probability * 100:.3f}%")
        return probability

from livekit.agents.voice import Agent
from livekit.plugins import openai, silero, gladia

class SimpleAgent(Agent):
    def __init__(self, userdata=None):
        self.custom_turn_detector = CustomTurnDetector(userdata)
        super().__init__(
            instructions="You are a helpful agent.",
            stt=gladia.STT(languages=["en", "fr", "de", "pt", "zh", "ja", "ko"]),
            llm=openai.LLM(model="gpt-4o-mini"),
            tts=openai.TTS(),
            vad=silero.VAD.load(),
            turn_detection=self.custom_turn_detector
        )

Advanced AudioRecognition Configuration:

from livekit.agents.voice.audio_recognition import AudioRecognition

audio_recognition = AudioRecognition(
    min_endpointing_delay=0.5,  # Minimum delay before responding
    max_endpointing_delay=6.0,  # Maximum delay for low confidence
    turn_detection_mode="vad"   # or "stt", "realtime_llm", "manual"
)

These configurations allow you to fine-tune the responsiveness and accuracy of your voice agent's turn-taking behavior based on your specific use case requirements. Sources: metrics/metrics_stt.py | livekit-plugins/livekit-plugins-turn-detector/livekit/plugins/turn_detector... | How does end-of-utterance detection work in conversations?