Inspect log Dec 19, 2025, 125941.778 PM Severity Logger livekit.plugins.silero Message inference is slower than realtime Attributes

{delay: 7.075579089000534

code.file.path: "/home/appuser/.local/lib/python3.11/site-packages/livekit/plugins/silero/vad.py"

code.function.name: "_main_task"

code.line.number: 397n

room_id: "RM_a2igfW2gunUk"

job_id: "AJ_pdRptcMZqe9v"

logger.name: "livekit.plugins.silero"

}

keep getting this error on livekit cloud service

Thanks for reaching out, Srikar. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

Im using livekit cloud why is this happening??

Thanks — since you’re using LiveKit Cloud, here’s why that warning appears and what you can try right away. Why this happens on LiveKit Cloud • The Silero VAD is taking longer to process audio than real time, so audio buffers build up and you see delays (your log showed ~7s). • Common Cloud causes: transient CPU contention or throttling on the worker running the VAD, high concurrency overall, or a temporary spike in load causing slower inference. Cold-start overhead can also cause occasional slow jobs. Immediate actions you can take • Tell me whether this happens for many rooms or just RM_a2igfW2gunUk, and if it’s constant or intermittent. That helps narrow whether it’s per-job or platform-wide. • If it’s intermittent and limited to few jobs: try reproducing at a quieter time to see if it’s load-related. • If it’s persistent across rooms/times: note the timestamps and rooms (you already have one), and share if you saw any recent increases in concurrent users or changes to your audio configuration. What I can do next for you • If you confirm it’s happening across multiple rooms or times, I’ll provide a short checklist of logs/metrics to collect and the specific next diagnostics to run so we can pinpoint whether this is platform resource contention, a model cold-start, or an unusual audio input. Which is it — limited to that one room/job or happening broadly and/or repeatedly?

literally only one room is open. Its strange. No other rooms were being run at the same time. Does livekit cloud use cpu or gpu?

Thanks — I’m responding to your question about why the Silero VAD warning appears when only one room is open, and whether LiveKit Cloud uses CPU or GPU. Short answer • LiveKit Cloud uses both CPU and GPU depending on the task; inference (STT/LLM/TTS/real-time voice models) will use GPU when applicable, while lighter VAD/turn-detection work can run on CPU. • The warning means the Silero VAD inference for that job took longer than real time (~7.08s in your log), so audio buffered and latency increased. Why this can happen with a single room • Temporary inference bottleneck (model loading, internal queuing, or a transient spike) can slow processing even with only one active room. • Some model pipelines include synchronous steps (STT → LLM → TTS) that add delay before VAD completes. • LiveKit Cloud manages hardware dynamically, so you don’t control instance type; occasional internal resource contention or model cold-starts can cause one-off slow jobs. Immediate troubleshooting steps • Was this a one-time event or repeated for that same room/job? If repeated, paste additional timestamps/jobs. • Reproduce test: try starting a short test call in the same region and room at a different time to see if delay persists. • Check whether your agent pipeline includes any synchronous heavy calls (e.g., large LLM requests) that run during the same job. If you want, I can: • Help interpret more logs if you paste additional entries around that timestamp. • Suggest specific diagnostics to capture (short list of logs/metrics) to pinpoint if this is a cold-start, queuing, or pipeline issue. Which would you prefer — paste more logs for me to inspect, or run a quick repro test and tell me whether it reoccurs? Sources: Deploy and scale agents on LiveKit Cloud | Building the all-in-one platform for voice AI agents | LiveKit turn detector plugin | LiveKit docs | Introducing LiveKit Inference: A unified model interface for voice AI | Troubleshooting Latency and Timeout Errors with Turn Detection on AWS