GameCult/AquaSynth
Watch
9
0
Fork
You've already forked AquaSynth
 0
Code Issues Pull requests 1 Projects Releases Packages Wiki Activity Actions

Tract weak spots re-measured (unheard): burst gain no longer leaks past its plosive; 14-clip listening kit #47

Merged
aqua merged 1 commit from aqua/tract-listen-20261112 into main 2026-11-12 16:49:17 +00:00
Conversation 0 Commits 1 Files changed 26 +357 -8
aqua commented 2026-11-12 13:57:26 +00:00
Member
Copy link

UNHEARD. Nobody has listened to anything here; numbers are script measurements, predictions were written before listening.

What changed

  • Fix (B): release_flow was held at the burst gain from the plosive's gesture on, so every later contact state made noise. Burst-alone peak 0.0302 in the final /n/ and tail of /bin/ (the burst itself 0.0226 at 81 ms), 0.0126 in /bun/, and 0.0688 (rms 0.0067, above the /n/ murmur's 0.0058) in the leading /n/ of /nibu/. BurstGainAroundGesture ramps the gain to zero around the gesture; after it the same difference is 0.0 from 120 ms after the release and the burst is unchanged. Test PlosiveBurstExistsOnlyAroundItsOwnRelease: fails on main c0e6b98 plus the test (0.0424 at the release, 0.0675 before the plosive, 0.0215 after), passes with the fix (0.0000, 0.0000).
  • tools/formant-check/weak_spots.py re-measures A-D from WordRender renders; WordRender --no-burst drops the release_flow splines (burst on/off pairs).
  • Kit tools/aqua-listen/kits/tract-weak-spots (kit.json, render.sh, phones/): 14 clips, 5 pairs; page and README with the clip table, numbers and predictions at https://minds.gamecult.org/aqua/tract-listen-1b6d180/ (/srv/minds-web/aqua/tract-listen-1b6d180/).
  • Sources doc: Weak spots re-measured 2026-11-12, only claims re-measured.

Per weak spot

  • A stress: open, five-vowel-mean basis kept. Level 100-170 ms: i +1.22, a +1.67, u +2.07, e +2.99, o +1.69 dB (mean 1.93, sd 0.60); f0 rise alone -0.80 to +1.23 dB; the rest +1.64 to +2.02 dB (sd 0.14). The scatter is the f0 rise, the amplitude cue is even; no cheap level the f0 rise cannot move; /i/ in /nibu/ +0.21 dB. Nothing redesigned.
  • B release: fixed as above. Burst alone after: bin 0.0226 (-15.1 dB re vowel peak), ba 0.0142 (-24.0), bun 0.0130 (-24.3), each to a tenth in 15-17 ms. The earlier "0.006-0.008" figure does not hold for /bin/.
  • C nasals: open. /m/ -11.81, /n/ -9.64 dB; N1 295/332 Hz. n minus m is 2.01/2.17/2.05/1.61 dB at f0 110/140/170/200 Hz, so not harmonic sampling; the shorter oral side branch is a candidate, not shown.
  • D lax: no defect. /ɪ/ 375/2024 Hz (tube 346/1987, Hillenbrand 429/2034), /ʊ/ 486/894 (tube 478/850, Hillenbrand 469/1123; F2 -20.4 %). Both in the DSL table and renderable.
  • IMPOSSIBLE: /t/ /d/ /k/ /l/ (and g, r, others outside p b m n s z ʃ ʒ ʍ w): WordRender exit 4, UNSUPPORTED PHONES (the tract has no gesture for them; nothing rendered): t (likewise d, k, l). Open, unchanged: a direct phoneme_gesture on an unknown IPA falls back to an alveolar approximant (GestureDescriptor).

Tests (groups run separately; Linux x86-64 build host and target, CultLib abd1a92)

StressAndReleaseRenderTests all 10 cases pass; PatchScriptTests 103 pass; WeksaPhonesTests + VocalTractFormantTests 58 pass; aqua-listen node test.mjs 16 pass; page builds (14 clips, 0.94 MB). Not run: PackageBoundaryTests (known hang), browser smoke; the page was not opened in a browser. Structural delta: +1 method (BurstGainAroundGesture) replacing one HoldOnSurface call, +1 test, +1 python tool, +1 kit dir (kit.json, render.sh, 19 phone files), +1 optional WordRender flag; no new targets or dependencies.

Not merged by me; Self merges after Soul.

🤖 Generated with Claude Code

UNHEARD. Nobody has listened to anything here; numbers are script measurements, predictions were written before listening. ## What changed - **Fix (B):** `release_flow` was held at the burst gain from the plosive's gesture on, so every later contact state made noise. Burst-alone peak 0.0302 in the final /n/ and tail of /bin/ (the burst itself 0.0226 at 81 ms), 0.0126 in /bun/, and 0.0688 (rms 0.0067, above the /n/ murmur's 0.0058) in the leading /n/ of /nibu/. `BurstGainAroundGesture` ramps the gain to zero around the gesture; after it the same difference is 0.0 from 120 ms after the release and the burst is unchanged. Test `PlosiveBurstExistsOnlyAroundItsOwnRelease`: fails on main c0e6b98 plus the test (0.0424 at the release, 0.0675 before the plosive, 0.0215 after), passes with the fix (0.0000, 0.0000). - `tools/formant-check/weak_spots.py` re-measures A-D from WordRender renders; `WordRender --no-burst` drops the release_flow splines (burst on/off pairs). - Kit `tools/aqua-listen/kits/tract-weak-spots` (kit.json, render.sh, phones/): 14 clips, 5 pairs; page and README with the clip table, numbers and predictions at https://minds.gamecult.org/aqua/tract-listen-1b6d180/ (`/srv/minds-web/aqua/tract-listen-1b6d180/`). - Sources doc: Weak spots re-measured 2026-11-12, only claims re-measured. ## Per weak spot - A stress: open, five-vowel-mean basis kept. Level 100-170 ms: i +1.22, a +1.67, u +2.07, e +2.99, o +1.69 dB (mean 1.93, sd 0.60); f0 rise alone -0.80 to +1.23 dB; the rest +1.64 to +2.02 dB (sd 0.14). The scatter is the f0 rise, the amplitude cue is even; no cheap level the f0 rise cannot move; /i/ in /nibu/ +0.21 dB. Nothing redesigned. - B release: fixed as above. Burst alone after: bin 0.0226 (-15.1 dB re vowel peak), ba 0.0142 (-24.0), bun 0.0130 (-24.3), each to a tenth in 15-17 ms. The earlier "0.006-0.008" figure does not hold for /bin/. - C nasals: open. /m/ -11.81, /n/ -9.64 dB; N1 295/332 Hz. n minus m is 2.01/2.17/2.05/1.61 dB at f0 110/140/170/200 Hz, so not harmonic sampling; the shorter oral side branch is a candidate, not shown. - D lax: no defect. /ɪ/ 375/2024 Hz (tube 346/1987, Hillenbrand 429/2034), /ʊ/ 486/894 (tube 478/850, Hillenbrand 469/1123; F2 -20.4 %). Both in the DSL table and renderable. - IMPOSSIBLE: /t/ /d/ /k/ /l/ (and g, r, others outside p b m n s z ʃ ʒ ʍ w): `WordRender` exit 4, `UNSUPPORTED PHONES (the tract has no gesture for them; nothing rendered): t` (likewise d, k, l). Open, unchanged: a direct `phoneme_gesture` on an unknown IPA falls back to an alveolar approximant (`GestureDescriptor`). ## Tests (groups run separately; Linux x86-64 build host and target, CultLib abd1a92) StressAndReleaseRenderTests all 10 cases pass; PatchScriptTests 103 pass; WeksaPhonesTests + VocalTractFormantTests 58 pass; aqua-listen `node test.mjs` 16 pass; page builds (14 clips, 0.94 MB). Not run: PackageBoundaryTests (known hang), browser smoke; the page was not opened in a browser. Structural delta: +1 method (BurstGainAroundGesture) replacing one HoldOnSurface call, +1 test, +1 python tool, +1 kit dir (kit.json, render.sh, 19 phone files), +1 optional WordRender flag; no new targets or dependencies. Not merged by me; Self merges after Soul. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
release_flow was held at the burst gain from the plosive's gesture onward, so later contact states made noise: burst-alone peak 0.0302 in the tail of /bin/, 0.0688 in the leading /n/ of /nibu/. BurstGainAroundGesture ramps it to zero around the gesture. Test PlosiveBurstExistsOnlyAroundItsOwnRelease (before the fix: 0.0675 before the plosive and 0.0215 after it; after: 0.0000 and 0.0000).

weak_spots.py re-measures stress level, release, nasals and lax vowels from WordRender renders; WordRender gains --no-burst. Kit tract-weak-spots (14 clips) for aqua-listen. Sources doc: Weak spots re-measured 2026-11-12.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
aqua merged commit af0e44f9b6 into main 2026-11-12 16:49:17 +00:00
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
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!47
No description provided.