Hear the difference

Voice architecture lab

Real local speech and Astra workers. Synthetic conversation context.

Supertonic + local Whisper

01 · Serial

Transcribe, finish the action work, then speak. The ORI scheduling baseline.

02 · Separate agents

Audio and action agents run concurrently; brief progress speech while work continues, completed results handed to speech.

03 · Shared state

Concurrent agents speak each merged state as results are committed; revisions refresh continuously.

Clara Martin

Continuity with the same buyer's text conversation

Ready

Headphones recommended. The microphone keeps listening during playback, so you can talk over the agent: the words you say become live steers on the turn that is already running, and the agent answers them without losing the work it finished. Silence ends your utterance; a turn that is cancelled is the only thing that discards work.

First audible chunk

—

First Astra content

—

Last transcript

No utterance yet

Live steer

Nothing steered yet

What actually happened

Live trace

Connect audio to begin. Timings appear from actual service events.

How to compare fairly

Every concurrent turn opens with a short non-semantic transport cue, then real Astra progress lines while the action agent works. The cue and the first Astra content are measured separately, so think time stays honest. Use the same buyer, prompt and audio. TTFS runs from received utterance through STT to the first PCM chunk; browser playback adds device latency. Text-triggered speech excludes STT. Model, cache and queue conditions affect the result.

The reproducible benchmark command records those conditions, completion latency, action evidence and interruptions for all three designs.

Recent conversation runs

Refresh measurements ↻

Architecture	Channel	Status	STT	Cue	First Astra content	Steers	Response complete
2 · Separate audio and action agents	voice	completed	—	92 ms	8379 ms	1	42200 ms
2 · Separate audio and action agents	voice	completed	1836 ms	1958 ms	8858 ms	3	55526 ms
2 · Separate audio and action agents	voice	completed	1776 ms	1873 ms	11456 ms	4	42909 ms
2 · Separate audio and action agents	voice	completed	1405 ms	1490 ms	8819 ms	3	54566 ms
2 · Separate audio and action agents	voice	completed	1425 ms	1524 ms	10346 ms	3	54771 ms
2 · Separate audio and action agents	voice	completed	1860 ms	1973 ms	9674 ms	3	51503 ms
2 · Separate audio and action agents	voice	completed	2200 ms	3566 ms	11335 ms	4	40724 ms
2 · Separate audio and action agents	voice	interrupted	1567 ms	1736 ms	—	0	—
2 · Separate audio and action agents	voice	completed	1858 ms	1985 ms	12274 ms	0	17980 ms
2 · Separate audio and action agents	voice	completed	1830 ms	1943 ms	19809 ms	4	36046 ms
2 · Separate audio and action agents	voice	completed	1315 ms	1407 ms	8740 ms	2	49568 ms
2 · Separate audio and action agents	voice	completed	1695 ms	1805 ms	19193 ms	4	46549 ms
2 · Separate audio and action agents	voice	interrupted	1378 ms	1553 ms	—	0	—
2 · Separate audio and action agents	voice	completed	1426 ms	1510 ms	18243 ms	4	42787 ms
2 · Separate audio and action agents	voice	completed	1240 ms	1450 ms	10304 ms	0	25266 ms
2 · Separate audio and action agents	voice	completed	2238 ms	2368 ms	22189 ms	5	46617 ms
2 · Separate audio and action agents	voice	failed	1705 ms	1864 ms	—	0	—
2 · Separate audio and action agents	voice	failed	1478 ms	1643 ms	—	0	—
2 · Separate audio and action agents	voice	failed	1782 ms	1905 ms	—	0	—
2 · Separate audio and action agents	voice	failed	1710 ms	1853 ms	—	0	—