Hi Team,
Wanted to flag a strange issue we’re noticing in the LiveKit voice agent pipeline — when the agent speaks, some completely unrelated background noise is audible in the audio. It’s not static or echo — it sounds like external sounds such as laughter, music, or other ambient noise that has no relation to the agent or the call.
This seems unusual and could indicate audio from another source bleeding into the agent’s audio track. Please help us here.
@Yethu_Krishnan, Which TTS or Realtime model is generating the agent audio (OpenAI Realtime, ElevenLabs, Cartesia, Gemini Live, etc.)?
Diagnosis for unrelated audio artifacts splits sharply by model, and a short clip uploaded directly to this thread (the Cloud session links are only accessible to LK staff) narrows it down between TTS hallucination, audio-processing artifact, and actual room-audio bleed.
Thank you for sharing your sessions as documented at Agent insights in LiveKit Cloud | LiveKit Documentation, I suggest you also un-share these once the issue is resolved.
I listened to your calls and it looks to me like the artefacts are on the USER’s audio, like this from the call ID ending 7qAh (it’s the gray waveform which corresponds to the user):
Is it possible this is audio from your user’s end?
I see you also have egress set up, do you also see these artefacts in your egress recording?
Those are very strange sounds, I’ve not heard them before.
This is from cartesia voice, and we are having this issue with one voice only.
I hear the artifacts you mention. Looks like you are using cartesia/sonic-3
Sonic-3 generates [laughter] by design when prompted with bracket tags. I assume that is not getting added? You might try dialing back the emotion temperature value in CartesiaOptions and see if that helps.
If I were debugging this, I would open a Cartesia ticket with the request_id to see what they find.
