Tract: ramps at phone boundaries, 12 ms stop release, smoothers starting at their input; Meta's tract-weak-spots-1 verdict and analysis #52

Merged
aqua merged 11 commits from aqua/tract-coarticulation into main 2026-11-13 22:54:18 +00:00
Member

Meta's tract-weak-spots-1 listening verdict (2 keep, 12 fix; data and analysis in docs/listening/) with every formant on target. Causes found by measuring the clips she heard: a step in the merged control curve at every phone boundary (nibu clicks, 0.40 and 0.49 peaks at 95 and 275 ms), si.smoo starting every control at 0 (each clip a 100 ms climb from silence), and a stop opening over 33 ms (a burst the length of a fricative).

Changes: 30 ms ramps at phone boundaries (PatchScript.WithPhoneTransition), a 12 ms stop release, smoo_init on control-surface sliders, a 12 ms vocal attack on the word scaffold, tools/formant-check/ear_metrics.py and phone_probe.py. Soul pass 1 found real regressions caused by this PR (closing-ramp hiss from the burst term, a voice bar 12 dB loud, smoo_init time constant 5 ms vs si.smoo 22.7 ms, stale Ding fixtures, stale 0.65 release windows, spline handles folded unlike the Faust lowering); fixed in 20aec42 and 3e54428. Soul pass 2 then found the velar test blind to leaks, /d/ closing noise untested and the smoo_from comment wrong; fixed in 96e6166..267a772 (the relaxed velar test is gone).

Nothing here has been heard by Meta. Every claim about sound is a measurement, not a listening.

Verified (head 7a72915; suite run at 267a772, one test host per class)

  • At 267a772 only three classes fail, all failing on main 71cf9ca too: FaustLoaderPlatformTests.FindLibraryLooksUnderLibForThePlatformName, ZynReferenceParityTests, SpeechDistributedTrainingTests. Every other class and every StressAndReleaseRenderTests case, run singly, exited 0 (s52/pr3/exits.txt in the Aqua workspace).
  • 7a72915 (two tests only) was then built and its two edited tests run alone: closing-noise 5/5 pass, velar closure 1/1 pass. The full suite was not rerun on 7a72915.
  • Soul pass 3 (aqua-body:verdict:cut-tract-coarticulation-pr52-fix2.s3, 3 Low findings, no Medium): the velar leak-injection test fails on every leak and closure-source mutant; the closing-noise test, source silent, fails on the burst-through-ramp mutants (p b t d k) with a -30 dB bound about 22 dB from each side; the smoo_from step form reproduces the old PlaceTable pins byte for byte; the glottal alias test passes at -43.4 dB and fails with the pitch smoother started at 0; freq/gain/note sliders still si.smoo from 0. Its Low findings: /g/ is not in TryIpaDescriptor so the closing-noise case could not fail (dropped in 7a72915); the velar seal bound was not the binding guard (now -15.5, 7a72915); the /k/ closure at or under /p/ criterion is unmet and has no ruling (follow-up). Not rechecked by a fourth Soul pass: 7a72915 is the two edits Soul prescribed, checked by the runs above.

Not verified

  • Nothing here has been heard. Phone recogniser (phone_probe.py) not re-run after the fixes (no torch on this host); weak_spots.py not run.
  • /k/ closure with the source on is louder than /p/ (-16.5 vs about -20 dB re vowels; it was on main too). Whether that is acceptable is unruled.

🤖 Generated with Claude Code

Meta's tract-weak-spots-1 listening verdict (2 keep, 12 fix; data and analysis in docs/listening/) with every formant on target. Causes found by measuring the clips she heard: a step in the merged control curve at every phone boundary (nibu clicks, 0.40 and 0.49 peaks at 95 and 275 ms), `si.smoo` starting every control at 0 (each clip a 100 ms climb from silence), and a stop opening over 33 ms (a burst the length of a fricative). Changes: 30 ms ramps at phone boundaries (`PatchScript.WithPhoneTransition`), a 12 ms stop release, `smoo_init` on control-surface sliders, a 12 ms vocal attack on the word scaffold, `tools/formant-check/ear_metrics.py` and `phone_probe.py`. Soul pass 1 found real regressions caused by this PR (closing-ramp hiss from the burst term, a voice bar 12 dB loud, `smoo_init` time constant 5 ms vs `si.smoo` 22.7 ms, stale Ding fixtures, stale 0.65 release windows, spline handles folded unlike the Faust lowering); fixed in 20aec42 and 3e54428. Soul pass 2 then found the velar test blind to leaks, /d/ closing noise untested and the `smoo_from` comment wrong; fixed in 96e6166..267a772 (the relaxed velar test is gone). **Nothing here has been heard by Meta.** Every claim about sound is a measurement, not a listening. ## Verified (head 7a72915; suite run at 267a772, one test host per class) - At 267a772 only three classes fail, all failing on main 71cf9ca too: `FaustLoaderPlatformTests.FindLibraryLooksUnderLibForThePlatformName`, `ZynReferenceParityTests`, `SpeechDistributedTrainingTests`. Every other class and every `StressAndReleaseRenderTests` case, run singly, exited 0 (`s52/pr3/exits.txt` in the Aqua workspace). - 7a72915 (two tests only) was then built and its two edited tests run alone: closing-noise 5/5 pass, velar closure 1/1 pass. The full suite was not rerun on 7a72915. - Soul pass 3 (`aqua-body:verdict:cut-tract-coarticulation-pr52-fix2.s3`, 3 Low findings, no Medium): the velar leak-injection test fails on every leak and closure-source mutant; the closing-noise test, source silent, fails on the burst-through-ramp mutants (p b t d k) with a -30 dB bound about 22 dB from each side; the `smoo_from` step form reproduces the old PlaceTable pins byte for byte; the glottal alias test passes at -43.4 dB and fails with the pitch smoother started at 0; freq/gain/note sliders still `si.smoo` from 0. Its Low findings: `/g/` is not in `TryIpaDescriptor` so the closing-noise case could not fail (dropped in 7a72915); the velar seal bound was not the binding guard (now -15.5, 7a72915); the `/k/` closure at or under `/p/` criterion is unmet and has no ruling (follow-up). Not rechecked by a fourth Soul pass: 7a72915 is the two edits Soul prescribed, checked by the runs above. ## Not verified - Nothing here has been heard. Phone recogniser (`phone_probe.py`) not re-run after the fixes (no torch on this host); `weak_spots.py` not run. - `/k/` closure with the source on is louder than `/p/` (-16.5 vs about -20 dB re vowels; it was on main too). Whether that is acceptable is unruled. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Meta's tract-weak-spots-1 verdict (2 keep, 12 fix) with formants matching their
targets: nibu clicks at the phone boundaries (a step in the merged control curve),
every clip a 100 ms climb from silence (si.smoo starts at 0), a 33 ms stop opening
heard as a fricative. ear_metrics.py and phone_probe.py measure those faults.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
smoo_init used a 5 ms pole where si.smoo is 1 - 44.1/SR (22.7 ms): live control moved 4.5x faster. The step is now
prev + (x - prev) * (1 - pole), so a constant control stays at its input to the bit. The source's own 10 ms pitch-ratio smoother
(si.smooth, from 0) rounded to ~4e-5 off its input once the control stopped climbing, which put the 2744 Hz alias-test source 0.12 Hz flat
and leaked its harmonics into the alias measurement (-30.3 dB, base -43.4, now -43.4); it is smoo_from at its input. The lowered dings gain
the two definitions only; no ding control uses them.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The exporter's burst noise is 4*opening*(1-opening) times a gain held through the whole stop gesture, so the 30 ms boundary ramp into a
stop was a 15-20 ms hiss (>4 kHz share of 6 ms frames at 188 ms of /ata/ -1 dB, base -57). The gain window now starts at
PatchScript.StopReleaseStartSeconds, the one owner of the release time (tests read it instead of 0.65 / 0.6 literals).
Stop lips are shut (0) rather than a 0.04 leak: word-initial /b/ voice bar -19.1 -> -27.7 dB, /p/ -30.1 -> -38.8 re the vowel, inside the
asserted -34..-26 and -60..-38. Spline handle times are a third of their segment: the Faust lowering ignores them, the C# timeline
folded a 12 ms opening with 23.75 ms handles. Closure windows start at the end of the boundary ramp (PhoneTransitionSeconds).
New tests: closing-ramp >4 kHz energy for /t/ /p/ /b/, the DSP's own stop opening read through its reservoir probe, the emitted smoother
(sample 0 and si.smoo pace) and the word's 12 ms attack rendered. Faust pins re-taken.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Soul s2 smoo-step-unfounded: from the input both forms hold a constant control to the bit, from 0 both stall 1e-5..5e-5 short of it;
the old form passes the alias test. The rewrite re-pinned every Faust id for nothing.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Soul s2 stale-release-literals.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Soul s2 release-split-owner (mutant m9: contact opening 20 ms early now fails 3 of 3) and nasal-lip-undefended (m14: nasal lips shut now fails /m/).
Read off the splines, no render.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Soul s2 velar-gap-hidden. The source-off closure cannot see a leak (m10/m11); the source-on closure after the ramp can. The 6.4 dB the closure
gained over main in the window after the ramp is the closing ramp itself (stepping the contact and constriction shut reads 6.5 dB lower); the
criterion 'at or under /p/' was unmet on main with the source on and is not restored.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Soul s2 closing-hf-no-d. With the source on /d/ -11.6 dB is voicing harmonics (-52.8 with the source off) and swings -13.5..-9.3 dB with f0 alone;
the burst gate held through the closure reads +5/+14 dB source-off and fails 5 of 6 (m4).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
With the step line swapped back the four old ids are reproduced exactly (p 784DFE4F27E0C6FA etc.), so that line is the whole delta.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
aqua merged commit ca6af1d678 into main 2026-11-13 22:54:18 +00:00
Sign in to join this conversation.
No reviewers
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
GameCult/AquaSynth!52
No description provided.