Skip to main content

Latency

Voice latency is the single biggest determinant of how a voice agent feels to a user. SimplAI emits per-turn telemetry from every voice run so you can monitor and improve it.

Per-turn metrics​

Every conversation turn emits these metrics:

  • TTFS (Time to First Speech) — time from user speech end to agent audio start. This is the headline metric: the user-perceived gap before the agent starts replying.
  • STT processing wait time — how long the transcriber took.
  • TTS generation wait time — how long the synthesizer took.
  • Queue wait time — only meaningful for Async mode, where requests pass through the dedicated voice priority queue.
  • Silence gap rate — count of silence gaps exceeding 1 second across the call.

Silence gaps are flagged when a delay between user speech end and agent audio start exceeds 1 second.

Spans in the trace​

All the metrics above appear as spans in the run detail trace — see Run history & traces for the trace view. STT and TTS spans include the transcribed/synthesized text as attributes, and they carry a provider_fallback_used: true flag when a fallback provider was used.

Aggregate dashboards​

The voice latency dashboard exposes percentile metrics:

  • P50 TTFS
  • P90 TTFS
  • P99 TTFS

Filters available:

  • Agent
  • Execution mode (Sync / Async / Realtime API)
  • Date

Use the dashboard to compare versions side-by-side after a configuration change, or to spot a regression after switching STT/TTS provider.

Alerting​

A built-in alert fires when P90 TTFS exceeds 1200ms for five consecutive turns on the same agent. Use this as a signal to:

  • Check whether a fallback STT or TTS provider has been triggered (look for provider_fallback_used: true spans in recent traces).
  • Check queue depth on Async-mode agents.
  • Check whether you've recently changed VAD settings or interruption duration.

Tuning checklist​

If you're trying to reduce TTFS:

  1. Switch to Sync mode if you're on Async and don't need queueing.
  2. Consider Realtime API mode for tool-light agents — it has the lowest TTFS of the three modes.
  3. Tune endpointing down — see Transcriber (STT).
  4. Tune VAD silence duration down — see Voice Activity Detection.
  5. Disable Use Speaker Boost in TTS settings — it improves clarity at the cost of latency.
  6. Reduce tool calls in the critical path — each Sync tool call adds its execution time to TTFS. Move non-critical work to post-call.
  7. Add fallback providers so a slow primary doesn't tank the whole turn.

Validate every change in Run history before declaring victory — TTFS is the truth.