Thanks — I’m looking at the initial LLM-response latency you described (greeting plays immediately, but the first LLM reply is slow while later replies stream fast). That behavior is common; here are the likely causes and the most effective mitigations. Likely causes • Cold start of VAD or local audio models: the VAD/STT model may be loading on the first connection, delaying turn detection and transcription. • LLM connection/model cold start: the agent may be establishing the LLM connection (DNS/TLS/auth) or loading model context on first request. • No preemptive/pipelined work on the first turn: prefetching, prewarming, or streaming pipelines typically kick in after the first interaction, so the first response is handled more synchronously. • Deployment cold start: if the agent process was scaled-to-zero, the container must boot and initialize before handling the request. • First STT/TTF (time-to-first-token) overhead from provider side: the initial STT or LLM request may incur extra latency before streaming begins. Recommended quick mitigations • Prewarm on startup: initialize VAD, STT/LLM clients and authenticate at agent startup so models/connections are ready before the first user speaks. • Keep a warm instance: run at least one always-on replica or use provisioned capacity so you avoid container cold starts. • Reuse persistent connections: ensure LLM/STT/TTS clients keep persistent sessions rather than creating new sockets per request. • Enable/verify streaming & preemptive generation: allow the agent to start generation earlier and stream tokens back as soon as they’re available. • Monitor TTFT and metrics: track Time-To-First-Token and related metrics to measure impact of each mitigation. If you want, tell me: • how your agent is deployed (serverless / always-on / autoscaled), • which STT/LLM providers you use, • and whether you already initialize clients in a startup/prewarm step— and I’ll give precise steps you can apply to your setup. Sources: LLM Metrics | LiveKit docs | LLM Output Replacement | LiveKit docs | Gemini Realtime Agent with Live Vision | LiveKit docs | Agent speech and audio | LiveKit docs | Deploying to production | LiveKit Docs