Thanks — I’m responding about your metrics increase (TTFT from ~0.4–0.5s to ~1s, and TTS TTFB from ~0.1s to ~0.8–0.9s) when using

session.on("metrics_collected")

. I know this is concerning — let’s narrow down causes and next steps. Quick context and targets • Typical targets: LLM TTFT <300ms, TTS TTFB <300ms, end‑to‑end <800ms. • Metric names vary by pipeline:

ttft

vs

llm_node_ttft

and

ttfb

vs

tts_node_ttfb

. Realtime-model pipelines emit different metrics (

RealtimeModelMetrics

). Possible reasons these values increased • Pipeline change: switching between standard STT→LLM→TTS and a realtime/unified model can change how metrics are measured. • Model/provider/version change: newer model versions or different providers can increase latency. • Network/region: increased network latency or a region change for services or your agent hosting. • Configuration/resource changes: CPU, memory, concurrency limits, or rate limiting on TTS/LLM nodes. • Instrumentation/metric changes: metric definitions or where you read them from may have changed (per-turn vs node-level). • Load/traffic: higher concurrent load can increase queuing and first-byte/token times. What I can help you do next • Compare per-turn metrics vs node-level metrics (e.g.,

conversation_item.metrics.llm_node_ttft

and

tts_node_ttfb

) to see if the delay is in a specific node. • Use the Agent Observability dashboard or Agent Console to view timeline-based breakdowns and synchronized audio playback to locate the bottleneck. • Verify whether you switched to a realtime model (then metrics emitted differ) or changed model/provider/region recently. Please provide the following so I can give specific diagnostic steps: 1. Any recent changes (model/provider/version, pipeline type, deployment region, or config)? 2. Are you comparing the same metric names (e.g., old

ttft

vs new

llm_node_ttft

) and using the same extraction path? 3. Are you seeing the increase consistently or only under load? Approx concurrent calls when you observe the higher numbers? 4. Whether you can share one example metric payload from

metrics_collected

or a

conversation_item_added

event (redact any PII). Give me those details and I’ll outline targeted checks and commands (or interpret a sample payload) to find whether this is measurement change vs real performance regression. Sources: How can I reduce latency in voice agents using STT, TTS and LLM? | Build Your First AI Voice Agent in Python: Complete Tutorial | LiveKit | Build Your First AI Voice Agent in Python: Complete Tutorial | LiveKit | LiveKit Agent Observability | Debug and optimize voice agents | LiveKit, with interactive elements