feat(vad): 512 ms silence tail (quiet mode 576 ms) so pauses don't split a sentence (#655)

[rosscado]

Why

The VAD closed a segment after 320 ms of silence, 2.4× shorter than Silero's own default. Spontaneous speech is full of 400–500 ms hesitations. In the #655 pause sweep, the old setting split a two-phrase sentence at a 400 ms pause 56% of the time, and at 500 ms 89% of the time. In production (saypi-api Stream A):

• a Pi turn averaged ~5 uploads;
• ~45% of Pi's sub-second clips were mid-sentence cuts.

Each fragment was transcribed without the rest of its sentence, often with no Whisper prompt at all, and scored for end-of-turn in isolation.

This is the second of the two #655 changes the founder approved. The first was the Silero v6 upgrade (#661).

What changes

preset	tail before	tail after
balanced (every host + dictation)	320 ms	512 ms
highSensitivity (quiet mode)	384 ms	576 ms (keeps its longer-than-balanced invariant)

Nothing else changes: thresholds, minimum speech and pre-pad are the same. The benchmark's latency column is now anchored to the pre-#655 320 ms tail, so its tables stay comparable, and the shipped row is labelled from the live config.

What it buys, and what it costs

AMI spontaneous-speech benchmark, Silero v6, npm run bench:vad:segmentation:

tail	thoughts split into 2+ uploads	clips per turn	short mid-turn cuts	extra latency on the turn's final upload
320 ms (before)	22%	2.47	78	–
512 ms (this PR)	16%	2.02	52	+192 ms
640 ms	11%	1.76	40	+320 ms

The cost is stated plainly: +192 ms from the end of speech to the turn's submit. The last clip uploads 192 ms later, and the submit timer counts from that upload, so the delay carries straight through. It doesn't compound. Dictation's live text also arrives in bigger chunks, since a pause now has to reach 512 ms. That's the knee of the curve; 640 ms buys a bit more for another +128 ms. Each clip also carries more trailing silence (about +250 ms of audio per turn), which matters if anything meters on audio duration.

Offsetting it (argued, not measured): about 0.45 fewer uploads per turn (fewer /transcribe round-trips and fewer Jev endpointing calls), and complete thoughts reaching both the ASR and the endpointing model.

I did not also anchor the submit timer to the true end of speech. That would shorten every patience-bound wait and change the endpointing behaviour the server eval is calibrated against, so it should be its own decision.

Measure after release: saypi-api's turn-outcome eval (#505) and Stream A can compare clips per turn, the ≤1 s share and premature-submit rate before and after.

Verification

• npm test: 3051 pass. The locked preset spec is updated, and the lifecycle spec now references VAD_CONFIGS instead of literals.
• Layer 3 Chrome: 48/48. The synthetic clips now carry ~1.2 s of trailing silence (was ~0.5 s). Chromium loops the fake-mic file, so at 0.5 s the new tail had zero margin at the loop seam; every clip now closes inside the file with ≥576 ms to spare (checked against the real v6 model).
• e2e-firefox VAD smoke: pass.

Refs #655

🤖 Generated with Claude Code

https://claude.ai/code/session_01W68zjaRTjQGYGCvoBA6aJ4

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-23T07:37:09.811602Z	a3e7095	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a3e7095d1a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Update the benchmark narrative for the new shipped tail

Changing the shipped tail to 512 ms makes the canonical benchmark write-up contradictory: bench/vad/README.md:250-285 still calls 320 ms the “shipped”/“today” configuration and describes latency relative to it, while the updated benchmark now labels 320 ms as pre-#655 and reports the shipped 512 ms policy at +192 ms. This can cause future tuning work to mistake the historical baseline for the live configuration, so update the fragmentation narrative and table labels alongside this value.

AGENTS.md reference: AGENTS.md:L104-L108

Useful? React with 👍 / 👎.

[rosscado]

Review: APPROVE WITH NITS (independent reviewer); nits addressed

• Latent flake (fixed in de20dda): the E2E fake-mic clips had ~0.5 s of trailing silence. Chromium loops the file, so a 512 ms tail had zero margin at the loop seam, and 576 ms never fired. The required e2e passed only by luck of the seam. I appended 0.7 s of digital silence to every clip (the speech is untouched) and set the generator to match. Verified with the real v6 model: every clip now closes as one segment with ≥576 ms to spare inside the file.
• Latency wording (fixed in the description): the +192 ms lands on the submit, not just the upload, because the timer counts from the VAD end event. It's still +192 ms and doesn't compound.
• Bench README "today" wording: fixed.
• Quiet mode 576 ms: forced by the existing invariant that quiet mode's tail exceeds balanced's (+64 ms, as before). Flagged to the founder for a one-line acknowledgement; not blocking.
• Verified by the reviewer:
  • Both values sit on the 32 ms grid (Math.floor(ms/32)).
  • No code in src/ assumes the old tail.
  • The Leverage the VAD's own confidence signal (computed, then discarded) to cut silence hallucinations — and characterize the thresholds #420 gate is unaffected: it's peak-based, so only telemetry means shift slightly.
  • The test edits are the deliberate lock update.
  • The bench anchoring is correct.