Thanks — I understand you disabled interruptions but user speech that happens while the agent is speaking still gets processed afterward. To drop any such audio when the agent is speaking, enable the session option that discards audio for uninterruptible speech. Use

discard_audio_if_uninterruptible=True

(Python) or

discardAudioIfUninterruptible: true

(TypeScript) when creating the AgentSession alongside

allow_interruptions=False

. Example: Python

session = AgentSession(
    llm=google.realtime.RealtimeModel(...),
    allow_interruptions=False,
    discard_audio_if_uninterruptible=True,
    # ...
)

TypeScript

const session = new voice.AgentSession({
  llm: new google.realtime.RealtimeModel(...),
  allowInterruptions: false,
  discardAudioIfUninterruptible: true,
  // ...
});

Would you like a snippet adapted to your existing session code or help confirming this setting is applied in your environment? Sources: Turns overview | LiveKit docs

Thanks — I’m responding to your new issue about longer Gemini realtime audio pausing mid-speech. Likely causes and quick checks • Network congestion / flow control: LiveKit may throttle media under packet loss or high jitter, which can cause audio stalls. Check network metrics (packet loss, jitter). • Client buffer underrun: Streaming TTS can exceed default buffer capacity for long responses — inspect client logs for buffer underflow or jitter buffer messages. • Playback interruptions: Ensure the audio renderer isn’t being re-mounted or otherwise interrupted during long playback. • Generation timing: Confirm

preemptive_generation

is disabled for realtime models (it can interfere with streaming). • SDK/server version mismatch: Older SDKs or server versions can show instability for longer streams. Immediate steps to try 1. Enable observability/insights and watch jitter, packet loss, and audio buffer levels. 2. Check client console logs for buffer underflow or WebRTC warnings. 3. Test with controlled network conditions (high/low bandwidth, simulated packet loss) to reproduce. 4. Try shorter TTS outputs to confirm length dependency. 5. Ensure you’re on the latest agent and client SDK versions and that

preemptive_generation

is off. If you’d like, I can dig deeper — please share: • Agent SDK version and client SDK + platform (e.g., @livekit/agents vX.Y, Chrome on desktop) • Typical audio duration when pauses occur (seconds) • Whether audio resumes after the pause or stops entirely • Any client/server logs showing warnings/errors Which of those details can you provide so I can help pinpoint the cause? Sources: Managing Video Quality During Network Congestion | OpenAI and LiveKit partner to turn Advanced Voice into an API | Agent speech and audio | LiveKit docs | examples/voice_agents | livekit/livekit_composite | DeepWiki | Text and transcriptions | LiveKit docs