Thanks — I understand the problem: turn detection is ending the user’s turn before the full number is spoken. That’s a common issue when capturing structured, sequential data like phone numbers. Here are practical steps you can try (ordered by easiest/most impactful) and a couple of clarifying questions so I can make more targeted suggestions. Quick checklist to try first • Update to the latest turn-detection model (MultilingualModel v0.4.1-intl). It’s specifically improved for structured inputs and reduces false interruptions. • Increase endpointing patience: ◦ Raise

min_endpointing_delay

(e.g., 0.8s) and

max_endpointing_delay

(e.g., 8s) so the detector waits longer before ending turns. • Make VAD less sensitive to very short pauses: ◦ Increase

min_silence_duration

(e.g., 0.8s) and/or lower activation sensitivity so brief pauses in spoken numbers aren’t interpreted as the end of speech. • If you use STT-based turn detection (e.g., AssemblyAI), increase its silence/timeouts (larger

min_turn_silence

and

max_turn_silence

) so the STT doesn’t cut the utterance early. Example configuration snippets 1. Use the improved multilingual turn detector:

from livekit.plugins.turn_detector.multilingual import MultilingualModel

session = AgentSession(
    turn_detection=MultilingualModel(),
    stt=your_stt_model,
    vad=your_vad,
    # ...
)

1. Increase endpointing delays:

session = AgentSession(
    turn_detection=MultilingualModel(),
    min_endpointing_delay=0.8,
    max_endpointing_delay=8.0,
    # ...
)

1. Make VAD tolerate brief pauses:

session = AgentSession(
    vad=silero.VAD.load(
        min_silence_duration=0.8,
        activation_threshold=0.4,
    ),
    # ...
)

When STT is used for turn detection (alternative):

session = AgentSession(
    turn_detection="stt",
    stt=assemblyai.STT(
        min_turn_silence=200,   # ms
        max_turn_silence=2000,  # ms
    ),
    min_endpointing_delay=0,
    # ...
)

Why this helps • The newer multilingual model recognizes structured formats (like phone numbers) and waits for completion instead of relying only on silence. • Increasing endpointing delays and VAD silence thresholds prevents short pauses between number groups (e.g., “(555) … 123 … 4567”) from being treated as end-of-turn. A few follow-up questions so I can refine recommendations • Which turn detector and model version are you currently using? • Are you using built-in VAD or a custom VAD (and which provider/version)? • Is STT performing turn detection, or are you using our turn-detector plugin? • Is the number format always the same (e.g., 10-digit US numbers) or does it vary? If you share the current model + VAD settings (or a minimal config snippet), I’ll suggest exact parameter values tuned to your calls. Sources: Improved End-of-Turn Model Cuts Voice AI Interruptions 39% | Agents Framework | livekit/livekit_composite | DeepWiki | VAD and turn detection configuration guide | LiveKit | AssemblyAI STT | LiveKit Documentation