VOICE SYSTEMS / INTERACTIVE WORKSHEET

Where does the pause go?

Budget the wait from the user’s last spoken sound to the first audible response. Start with a hypothetical serial path, then replace the inputs with measurements from your own traces.

Six stages. One explicit assumption.

All stages below run in sequence in this model. Do not enter overlapping durations twice. Every default is illustrative—not a recommended target or measured product result.

01 / End-of-turn detectionFrom the last spoken sound until the service decides the user finished.ms02 / Remaining transcriptionOnly transcription still on the critical path after end-of-turn detection.ms03 / Evidence retrievalRequired lookup before generation begins in this serial example.ms04 / First usable textUntil enough generated text is available to start speech.ms05 / First audio generationUntil the speech service produces the first usable audio chunk.ms06 / Delivery and playoutDelivery, buffering and device delay before audible playback.msYour illustrative target0–30,000 ms. A comparison hypothesis, not a service-level agreement.ms

SERIAL CRITICAL PATH / NOT A BENCHMARK

1,330milliseconds to first audio

670 ms below your illustrative target. Largest stage: First usable text (400 ms).

End-of-turn detection350 ms
Remaining transcription180 ms
Evidence retrieval120 ms
First usable text400 ms
First audio generation200 ms
Delivery and playout80 ms

350 + 180 + 120 + 400 + 200 + 80 = 1,330 ms

Learn to replace assumptions with a trace ↗

The boundary of the model

The sum includes only the serial durations you enter. Real systems may transcribe while a person speaks, retrieve speculatively, overlap text and speech, or introduce queueing and network delays. Model the actual critical path before interpreting a total. This worksheet does not predict a percentile, full-response duration, interruption responsiveness, lip synchronization or user satisfaction.

A lower total does not prove better turn-taking: overly aggressive end-of-turn detection can cut the user off. Compare traces with speech-repair outcomes and turn ownership, and inspect timeout budgets separately. Inputs are local to this page and are not stored or transmitted.

Original editorial model, prepared 19 September 2026. No live model, microphone or remote API is used.