voice-source-1: glottal pulse and noise-colour listening kit (unheard) #53

Merged
aqua merged 12 commits from aqua/voice-source-kit into main 2026-11-14 00:36:35 +00:00
Member

voice-source-1: a 10-clip listening kit (5 pairs of the phrase /ɑmi/ at 140 Hz, each pair differing in one voice-source field) for Meta's open question about why the kept /i/ clips sound like a kazoo. Entry points: tools/aqua-listen/kits/voice-source/ (predictions.md committed before any render, render.sh, clips.py, hnr.py, kit.json), WordRender --source, FaustExport.RosenbergPulse, source fields pulse and noise_hp on source_port.

Nothing here has been heard. Every number is a measurement; no listening verdict exists for any clip. Published page (unheard): https://minds.gamecult.org/aqua/voice-source-1-e343d2b/listen.html, rendered on main ca6af1d (PR #52). Two earlier pages (voice-source-1-3800414, -51084c3) were withdrawn: their titles carried predictions and they shared an answer key.

Known and stated, not fixed:

  • The Rosenberg band-energy figures (now in predictions.md, not on the page) are from the render before PR #52. A different FFT estimate on the new render differs by under 1 dB, but the original method could not be reproduced, so they were not re-taken.
  • The Rosenberg render is 2.5 dB lower in output rms and 6.9 dB in peak than the saw (pair 2 matches the level, gain 1.341). LF pulse is not tested; one phrase, one f0.
  • phone_probe.py was not run (no torch on this host).

Soul pass 1 (aqua-body:verdict:cut-voice-source-kit.s1): the five promises held (saw default byte-identical to main, one field per pair, figures true, predictions committed before any render, nothing claimed heard); 4 Medium and 4 Low findings, none above Medium. Fixed by Hands (report ...cut-voice-source-kit.h2, a29cad3..e343d2b): waveform test for the Rosenberg pulse that kills Soul's three mutants, --source refuses typos and malformed input with exit 2, noise_hp corner pinned, neutral clip titles (predictions and measured table only in predictions.md), written record corrected, bytecode ignored. The withdrawn pages cover the answer-key finding.

Verified at e343d2b: GlottalSourceSpectrumTests 9, PatchScriptTests 104, PlaceTableTests 7, DingCatalogTests 10, NasalCouplingTests 7, StopClosureAndControlRenderTests 9, WeksaPhonesTests 32, PhoneTransitionTests 5, all passing; built page contains no PREDICTION/FALSIFIED. NOT run: the full per-class suite at e343d2b (at cc64607 Soul ran 283/286 with three failures that also fail on main). No second Soul pass over the fixes: the first found nothing above Medium and the fixes are the ones it prescribed, checked by the runs above.
Not fixed here: the tract-weak-spots and tract-stops kit titles still carry predictions (follow-up pr48-soul-lows-20261113).

🤖 Generated with Claude Code

voice-source-1: a 10-clip listening kit (5 pairs of the phrase /ɑmi/ at 140 Hz, each pair differing in one voice-source field) for Meta's open question about why the kept /i/ clips sound like a kazoo. Entry points: `tools/aqua-listen/kits/voice-source/` (`predictions.md` committed before any render, `render.sh`, `clips.py`, `hnr.py`, `kit.json`), `WordRender --source`, `FaustExport.RosenbergPulse`, source fields `pulse` and `noise_hp` on `source_port`. **Nothing here has been heard.** Every number is a measurement; no listening verdict exists for any clip. Published page (unheard): https://minds.gamecult.org/aqua/voice-source-1-e343d2b/listen.html, rendered on main `ca6af1d` (PR #52). Two earlier pages (voice-source-1-3800414, -51084c3) were withdrawn: their titles carried predictions and they shared an answer key. Known and stated, not fixed: - The Rosenberg band-energy figures (now in predictions.md, not on the page) are from the render before PR #52. A different FFT estimate on the new render differs by under 1 dB, but the original method could not be reproduced, so they were not re-taken. - The Rosenberg render is 2.5 dB lower in output rms and 6.9 dB in peak than the saw (pair 2 matches the level, gain 1.341). LF pulse is not tested; one phrase, one f0. - `phone_probe.py` was not run (no torch on this host). Soul pass 1 (`aqua-body:verdict:cut-voice-source-kit.s1`): the five promises held (saw default byte-identical to main, one field per pair, figures true, predictions committed before any render, nothing claimed heard); 4 Medium and 4 Low findings, none above Medium. Fixed by Hands (report `...cut-voice-source-kit.h2`, a29cad3..e343d2b): waveform test for the Rosenberg pulse that kills Soul's three mutants, `--source` refuses typos and malformed input with exit 2, noise_hp corner pinned, neutral clip titles (predictions and measured table only in predictions.md), written record corrected, bytecode ignored. The withdrawn pages cover the answer-key finding. Verified at e343d2b: GlottalSourceSpectrumTests 9, PatchScriptTests 104, PlaceTableTests 7, DingCatalogTests 10, NasalCouplingTests 7, StopClosureAndControlRenderTests 9, WeksaPhonesTests 32, PhoneTransitionTests 5, all passing; built page contains no PREDICTION/FALSIFIED. NOT run: the full per-class suite at e343d2b (at cc64607 Soul ran 283/286 with three failures that also fail on main). No second Soul pass over the fixes: the first found nothing above Medium and the fixes are the ones it prescribed, checked by the runs above. Not fixed here: the tract-weak-spots and tract-stops kit titles still carry predictions (follow-up pr48-soul-lows-20261113). 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
aqua merged commit a9e2ce5841 into main 2026-11-14 00:36:35 +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!53
No description provided.