Available for LiveKit Agents, Pipecat, and any custom stack via the OTLP HTTP exporter. See Turn on traces.
Where traces fit
Voice Metrics tells you which turn was slow and which stage dominated it, on every call and with no setup. A trace tells you what ran inside that stage: the model request, the streaming adapter, each tool call. Use it when the stage number does not explain the delay. An LLM stage that was really a slow tool call, or a TTS stage that was slow in the stream adapter rather than at the provider, look identical in a stage breakdown and obvious in a span tree.Reading the span tree
Tuner renders the trace against a time axis, with each bar positioned relative to the start of the session.
- Follow the widest bar at each level. Above,
agent_turnis 4.84s and thetts_nodeunder it is 4.62s, so synthesis dominated that turn, not the model. - Siblings overlap.
llm_nodeandtts_nodeboth start at 00:00, because synthesis streams on the first tokens. Children do not sum to their parent, the same way the pipeline stages do not sum to TTFB. - Collapse large traces. The call above has 171 spans. Collapse to the turn level, then expand the one turn you care about. Click a span to open its details.
Span names come from your framework. See LiveKit traces or Pipecat traces.
Turn on traces
Setup lives with your integration.LiveKit traces
Send traces from LiveKit Agents.
Pipecat traces
Send traces from Pipecat.
Custom stack traces
Send traces from any OTLP-compatible stack via the Direct API.
What Tuner stores
Traces are separate from the call record, which is still governed by PII Redaction and Data Retention.
Next steps
Voice metrics and latency
The per-turn stage breakdown you start from before opening a trace.
Root cause analysis
Working through a single failing call end to end.