Latency

Four numbers, and only one of them is felt.

T_first	time to the first audible thing
T_useful	time to something the caller can act on
T_complete	total
T_perceived	what they would say afterwards

Almost every optimization attacks T_complete. Callers feel T_first and remember T_perceived, which tracks silence and variance, not duration. A twelve-second answer narrated every three seconds feels responsive; four seconds of nothing feels broken.

The budget

Estimates, not measurements. The panel replaces each one with a measured number, and this table is rewritten when it does.

Segment	Estimate
WebSocket in	20–40 ms
End of utterance	500–800 ms fixed; 200–400 semantic, on a GPU-served model
ASR finalize	50–150 ms
Host TTFT	400–700 ms uncached; 200–400 once the cache patch lands
TTS time-to-first-byte	80–200 ms; ≈ 0 pre-rendered
Playback	40–80 ms

Tier	First audio	Complete
conversational	600–900 ms	same
definitional	600–900 ms	1.4–1.9 s
analytical	500–900 ms	8–14 s p50

First audio is tier-independent, and that is the whole trick. The host speaks before any tool resolves, so a caller cannot tell from the onset which tier they hit.

The switches

Each is a setting the graph reads and a delta the panel shows. A lever an audience cannot hear is a claim, not a demonstration.

#	Setting	Off	On
1	DAV_EOU_MODE	fixed silence (CI)	the turn detector, on a GPU
2	DAV_CACHE_PREFIX	no breakpoint	cached prefix, once the patch lands
3	DAV_HOST_MODEL	Opus	Haiku
4	DAV_TTS_CHUNK	wait for the whole response	speak each sentence
5	DAV_PRERENDERED_ACK	synthesize live	served from disk
6	DAV_SPECULATIVE_DISPATCH	wait for a final transcript	guess on a confident partial

Defaults are the CI shape, because a default nobody exercises is not a default.

Where the seconds actually are

Worth stating plainly, because it is easy to spend effort in the wrong place:

Lever	Effect on the analytical path	Whose
Model and effort per hop	−8 to −12 s	upstream’s
Prompt caching	−1.5 to −4 s	upstream’s
Parallel tool calls	7 hops → 4–5	upstream’s
Sub-agent fan-out	−1 to −3 s	upstream’s
Everything in this repository	0 s	ours

Nothing here makes the agent faster. What it changes is T_first and T_perceived — from twenty-six seconds of silence to under a second of acknowledgement, followed by narration. That is the entire contribution, and it is worth more than any of the rows above.

Two things that are easy to get wrong

Speaking per sentence beats every token-level optimization. The caller hears the first clause, not the last token. Total generation time is nearly irrelevant to how it feels.

Tokens not generated are the cheapest latency win there is. The answer is spoken as headline, then definition, then caveats — so the useful part arrives first and the rest can be interrupted without loss.