I'm checking out the GPT Realtime model integration (gpt-realtime-2025-08-28) and all calls I place stop working after around 2:30min -- the agent doesn't detect any normal responses anymore and then hangs up after the silence timeout. Agent setup: { "id": "d7eeed91-3280-48a5-ba50-242de3eefc46", "voice": { "model": "tts-1", "voiceId": "marin", "provider": "openai" }, "model": { "model": "gpt-realtime-2025-08-28", "provider": "openai", "_systemPromptChars": 27316 }, "transcriber": { "model": "nova-3", "language": "en-US", "provider": "deepgram", "smartFormat": true }, "firstMessage": "Hey, can you hear me okay?", "firstMessageInterruptionsEnabled": false, "silenceTimeoutSeconds": 30, "maxDurationSeconds": 900, "backgroundDenoisingEnabled": true, "backgroundSound": "office", "startSpeakingPlan": { "waitSeconds": 0.4, "smartEndpointingPlan": { "provider": "livekit" } }, "stopSpeakingPlan": { "numWords": 3, "voiceSeconds": 0.4, "backoffSeconds": 1 }, "endCallFunctionEnabled": true }

Hey, I’ve worked with realtime voice agents and this sounds like a session/stream timeout or audio pipeline stall around the 2–3 min mark. I can help you debug it quickly (LiveKit endpointing, Deepgram stream, and Realtime model session handling) and get it stable for full call duration. @btt