voice-source-1: neutral clip titles; measured figures, predictions and falsifiers live only in predictions.md; noise .05 and the sixteen calibration renders recorded; README no longer says titles are hidden
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
This commit is contained in:
parent
1c18ca5347
commit
e343d2bc4c
3 changed files with 35 additions and 14 deletions
|
|
@ -39,7 +39,7 @@ node smoke.mjs dist/listen.html # optional
|
|||
```
|
||||
|
||||
- `id`, `file` (wav, mp3, ogg or flac, relative to `--clips`), `title`, `feel` (intended feel; shown to the listener), `group`: required. `pair`: optional, must name another clip whose `pair` names this one. Anything wrong stops the build with the full list of problems.
|
||||
- Title, id and group are hidden from the listener and appear only in the export. The intended feel is shown, since "did it land" needs it.
|
||||
- Title, id and group are not shown beside a clip, but they are in the page's bytes and in the export, which a listener can press at any time: they are not hidden. Keep a title a short neutral label; nothing a listener should not read goes in a title. The intended feel is shown, since "did it land" needs it.
|
||||
- Order is shuffled from a random seed kept with the listener's answers (it is in the export). A pair is shuffled as one unit, with A/B position shuffled inside it.
|
||||
- Audio is embedded as-is (base64), so the page is about 1.4x the clip bytes. The build warns over 100 MB.
|
||||
- Autosave key is the kit id plus the sorted clip ids (`aqua-listen.v2:<kit id>:<ids>`), not the content hash: a rebuilt page (edited note, re-rendered clip) keeps the listener's answers, and a kit with a different clip-id set starts fresh while the old entry stays in the browser. Saved state is the shuffle seed, answers by clip id and pair choices by `idA|idB`; a pair's A/B/same value is positional (A is the first of the pair in the seed's order), so it holds only while the manifest's clip order is unchanged. When the new key is empty, the page adopts the fullest earlier `aqua-listen:<kit id>:<hash>` save whose answers name only this kit's clips; a save already under the new key is never overwritten. The save also records `pairSig`, the manifest's clip ids in manifest order: on load, a save whose `pairSig` is missing or different keeps its seed and answers and drops its pair picks, so reordering or re-pairing clips never reassigns a pick (an adopted hash-keyed save has no signature and cannot be checked, so its pairs are kept and stamped with the current one). A save missing `pairs`, or with `answers` that is not an object, is normalised to empty rather than crashing the page. "Erase this kit's saved answers" removes the new key and any adoptable hash-keyed save of this kit.
|
||||
|
|
@ -58,7 +58,7 @@ The blank verdict files this was meant to match were not found under `/opt/commo
|
|||
### Kits in the repo
|
||||
|
||||
`kits/voice-source` is the voice-source kit: one phrase, five pairs, each pair changing one source parameter (`pulse=`, `noise=`, `noise_hp=` on the scaffold's `source_port`, set through `WordRender --source`); `predictions.md` was committed before any clip was rendered, `hnr.py` names the noise settings in dB, `clips.py` makes the clip files. `kits/tract-weak-spots` and `kits/tract-stops` hold a `kit.json`, the `phones/*.json` that WordRender reads, and a `render.sh` that renders the clips (`render.sh <repo-root> <out-root>`; needs CULTLIB_ROOT and AQUASYNTH_FAUST_HOME, see the script header). Build a page from the rendered clips:
|
||||
`node tools/aqua-listen/build.mjs --manifest tools/aqua-listen/kits/tract-stops/kit.json --clips <out-root>/clips --out <out-root>/dist`. The titles in those kit.json files carry measured figures, a prediction and a falsifier; the page hides them from the listener.
|
||||
`node tools/aqua-listen/build.mjs --manifest tools/aqua-listen/kits/tract-stops/kit.json --clips <out-root>/clips --out <out-root>/dist`. The titles in `kits/tract-weak-spots` and `kits/tract-stops` still carry measured figures, a prediction and a falsifier, and the page does not hide them (they are in its bytes and in the export), so those pages are not blind to them. The `kits/voice-source` titles are neutral condition labels; its figures, predictions and falsifiers are in `predictions.md`.
|
||||
|
||||
### Pictures and burst notes (for a reader who cannot listen)
|
||||
|
||||
|
|
|
|||
|
|
@ -10,7 +10,7 @@
|
|||
"file": "saw_a.wav",
|
||||
"group": "pulse",
|
||||
"pair": "pulse-rosenberg",
|
||||
"title": "Measured: HNR 30.3 dB, output rms 0.0537, peak 0.227. Scaffold source: band-limited saw through the tilt lowpass.",
|
||||
"title": "saw pulse, scaffold noise",
|
||||
"feel": "the short made-up phrase 'ah-mee' in a synthetic voice; listen to the voice itself (buzz, edge, breath), not the words"
|
||||
},
|
||||
{
|
||||
|
|
@ -18,7 +18,7 @@
|
|||
"file": "ros.wav",
|
||||
"group": "pulse",
|
||||
"pair": "pulse-saw",
|
||||
"title": "Measured: HNR 29.5 dB, output rms 0.0401 (-2.5 dB re saw), peak 0.102 (-6.9 dB). Band energy re total, 0-0.5/0.5-1.5/1.5-3/3-6 kHz, from the render before the tract fix (a different FFT estimate on the new render differs by under 1 dB; not re-measured with the original method): saw -2.3/-4.0/-24.0/-24.9 dB, Rosenberg -0.3/-12.4/-43.7/-40.3 dB. PREDICTION: darker and rounder than the saw, less buzz and edge, F2 of /i/ weaker; loudness the same (that last part is already falsified at the output: it is 2.5 dB quieter). FALSIFIED if no difference is heard, if it is heard as buzzier or brighter, or if it is heard as the same kazoo quality.",
|
||||
"title": "Rosenberg pulse, scaffold noise",
|
||||
"feel": "the short made-up phrase 'ah-mee' in a synthetic voice; listen to the voice itself (buzz, edge, breath), not the words"
|
||||
},
|
||||
{
|
||||
|
|
@ -26,7 +26,7 @@
|
|||
"file": "saw_b.wav",
|
||||
"group": "pulsematch",
|
||||
"pair": "pulsematch-rosenberg",
|
||||
"title": "Measured: HNR 30.3 dB, output rms 0.0537, peak 0.227. Byte copy of pulse-saw's render.",
|
||||
"title": "saw pulse, scaffold noise",
|
||||
"feel": "the short made-up phrase 'ah-mee' in a synthetic voice; listen to the voice itself (buzz, edge, breath), not the words"
|
||||
},
|
||||
{
|
||||
|
|
@ -34,7 +34,7 @@
|
|||
"file": "ros_matched.wav",
|
||||
"group": "pulsematch",
|
||||
"pair": "pulsematch-saw",
|
||||
"title": "Measured: HNR 29.5 dB, output rms 0.0537 (matched to the saw), peak 0.137. The Rosenberg render times 1.341 (+2.55 dB). PREDICTION: the same darker, rounder voice as in pair 'pulse', and no longer the quieter one. FALSIFIED if at matched rms it is no longer heard as darker/rounder (then loudness carried the first pair), or if it is heard as louder or brighter.",
|
||||
"title": "Rosenberg pulse, level matched to the saw",
|
||||
"feel": "the short made-up phrase 'ah-mee' in a synthetic voice; listen to the voice itself (buzz, edge, breath), not the words"
|
||||
},
|
||||
{
|
||||
|
|
@ -42,7 +42,7 @@
|
|||
"file": "clean.wav",
|
||||
"group": "noise",
|
||||
"pair": "noise-hnr20",
|
||||
"title": "Measured: HNR 30.3 dB, output rms 0.0537, peak 0.227. The scaffold's noise floor (0.05); noise 0 measured 30.8 before the tract fix, the method's ceiling here.",
|
||||
"title": "scaffold noise",
|
||||
"feel": "the short made-up phrase 'ah-mee' in a synthetic voice; listen to the voice itself (buzz, edge, breath), not the words"
|
||||
},
|
||||
{
|
||||
|
|
@ -50,7 +50,7 @@
|
|||
"file": "w20.wav",
|
||||
"group": "noise",
|
||||
"pair": "noise-clean",
|
||||
"title": "Measured: HNR 19.9 dB, output rms 0.0539, peak 0.242. White glottal noise raised to 0.33. PREDICTION: audible breath/hiss under the voice; a listener who called the clean voice a kazoo or buzzer hears this as less pure-tone but still a machine. FALSIFIED if B is not noisier; if the noise is heard as a separate hiss stream not fused with the voice; or if A and B are the same kazoo to the same degree (then purity of the periodicity is not what makes it a kazoo).",
|
||||
"title": "white noise, level .33",
|
||||
"feel": "the short made-up phrase 'ah-mee' in a synthetic voice; listen to the voice itself (buzz, edge, breath), not the words"
|
||||
},
|
||||
{
|
||||
|
|
@ -58,7 +58,7 @@
|
|||
"file": "w20_b.wav",
|
||||
"group": "colour",
|
||||
"pair": "colour-hp20",
|
||||
"title": "Measured: HNR 19.9 dB, output rms 0.0539, peak 0.242. Byte copy of noise-hnr20's render.",
|
||||
"title": "white noise, level .33",
|
||||
"feel": "the short made-up phrase 'ah-mee' in a synthetic voice; listen to the voice itself (buzz, edge, breath), not the words"
|
||||
},
|
||||
{
|
||||
|
|
@ -66,7 +66,7 @@
|
|||
"file": "h20.wav",
|
||||
"group": "colour",
|
||||
"pair": "colour-white20",
|
||||
"title": "Measured: HNR 19.9 dB, output rms 0.0540, peak 0.238. Noise high-passed at 2 kHz (2nd order), level 0.33. PREDICTION: reads as hiss or 's'-like air, more separable from the voice than the white noise at the same HNR; the white one reads as roughness or breathy voice. FALSIFIED if no difference is heard, or if B is less hissy than A.",
|
||||
"title": "noise high-passed at 2 kHz, level .33",
|
||||
"feel": "the short made-up phrase 'ah-mee' in a synthetic voice; listen to the voice itself (buzz, edge, breath), not the words"
|
||||
},
|
||||
{
|
||||
|
|
@ -74,7 +74,7 @@
|
|||
"file": "h25.wav",
|
||||
"group": "aspiration",
|
||||
"pair": "aspiration-hp15",
|
||||
"title": "Measured: HNR 25.8 dB, output rms 0.0538, peak 0.231. High-passed noise at level 0.165. PREDICTION: at most faintly airy, may be called no different from a clean voice. FALSIFIED if A is already clearly hissy.",
|
||||
"title": "noise high-passed at 2 kHz, level .165",
|
||||
"feel": "the short made-up phrase 'ah-mee' in a synthetic voice; listen to the voice itself (buzz, edge, breath), not the words"
|
||||
},
|
||||
{
|
||||
|
|
@ -82,7 +82,7 @@
|
|||
"file": "h15.wav",
|
||||
"group": "aspiration",
|
||||
"pair": "aspiration-hp25",
|
||||
"title": "Measured: HNR 14.9 dB, output rms 0.0546, peak 0.250. High-passed noise at level 0.6. PREDICTION: clearly hissy, breathier than A. FALSIFIED if B is not hissier than A.",
|
||||
"title": "noise high-passed at 2 kHz, level .6",
|
||||
"feel": "the short made-up phrase 'ah-mee' in a synthetic voice; listen to the voice itself (buzz, edge, breath), not the words"
|
||||
}
|
||||
]
|
||||
|
|
|
|||
|
|
@ -7,7 +7,7 @@ Source-side numbers (computed from the two source formulas alone, harmonics of 1
|
|||
| pair | A | B | what changes | prediction | falsified if |
|
||||
|---|---|---|---|---|---|
|
||||
| pulse | saw | rosenberg | `pulse=` | B is darker and rounder: less buzz, less edge, F2 of /i/ weaker (the harmonics near 2 kHz are ~22 dB lower). Loudness the same (rms within 0.5 dB). | no difference heard; or B heard as buzzier/brighter; or B heard as the same "kazoo" quality. |
|
||||
| noise level | saw, noise at scaffold (.04) | saw, white noise raised to HNR 20 dB | `noise=` | B has audible breath/hiss under the voice; a listener who called A a "kazoo" or "buzzer" hears B as less pure-tone, but still a machine. | B is not noisier; or the noise is heard as a separate hiss stream not fused with the voice; or A and B are both still a kazoo to the same degree (then purity of the periodicity is not what makes it a kazoo). |
|
||||
| noise level | saw, noise at scaffold (effective level .05: the gesture splines hold noise at .05 and win over the slider's base value) | saw, white noise raised to HNR 20 dB | `noise=` | B has audible breath/hiss under the voice; a listener who called A a "kazoo" or "buzzer" hears B as less pure-tone, but still a machine. | B is not noisier; or the noise is heard as a separate hiss stream not fused with the voice; or A and B are both still a kazoo to the same degree (then purity of the periodicity is not what makes it a kazoo). |
|
||||
| noise colour | white noise, HNR 20 dB | noise high-passed at 2 kHz, HNR 20 dB | `noise_hp=` (noise level re-set to keep HNR 20) | B reads as hiss/"s"-like air, more separable from the voice than A; A reads as roughness/breathy voice. | no difference; or B heard as less hissy than A. |
|
||||
| aspiration level | high-passed noise, HNR 25 dB | high-passed noise, HNR 15 dB | `noise=` | B is clearly hissy; A is at most faintly airy and may be called "no difference" from a clean voice. | A is already clearly hissy, or B is not hissier than A. |
|
||||
|
||||
|
|
@ -17,8 +17,29 @@ What this kit cannot say: LF (the Liljencrants-Fant model) is not tested; the pu
|
|||
|
||||
## Added after the calibration renders, before the kit renders
|
||||
|
||||
Eight calibration renders of the same phrase were made to turn noise settings into dB (nobody listened to them). They gave: noise 0 -> HNR 30.8 dB (the method's ceiling on this phrase), white noise .1/.16/.25 -> 27.7/25.2/22.2, noise high-passed at 2 kHz .2/.4/.7/1.0 -> 24.5/18.4/13.6/10.6, and Rosenberg at the scaffold's noise -> 29.1 dB. They also showed the Rosenberg render is 2.4 dB lower in output rms and 6.8 dB lower in peak than the saw, although the SOURCE rms is matched: the tract passes the saw's strong upper harmonics and the pulse has none. The first prediction row's "loudness the same" is therefore already known to be false at the output; it stays in the table as written.
|
||||
Sixteen calibration renders of the same phrase were made in two rounds of eight (14:32 and 14:43 on 2026-11-13; nobody listened to them). The figures below are from the second round plus the Rosenberg render of the first; the first round's other five renders were made and their results are not reported here. They gave: noise 0 -> HNR 30.8 dB (the method's ceiling on this phrase), white noise .1/.16/.25 -> 27.7/25.2/22.2, noise high-passed at 2 kHz .2/.4/.7/1.0 -> 24.5/18.4/13.6/10.6, and Rosenberg at the scaffold's noise -> 29.1 dB. They also showed the Rosenberg render is 2.5 dB lower in output rms (measured -2.54 dB on the final render) and 6.8 dB lower in peak than the saw, although the SOURCE rms is matched: the tract passes the saw's strong upper harmonics and the pulse has none. The first prediction row's "loudness the same" is therefore already known to be false at the output; it stays in the table as written.
|
||||
|
||||
A fifth pair is added so loudness cannot carry the answer: the Rosenberg render scaled up by its measured rms deficit against the saw (a gain on the clip, applied by the kit's render script), against the same saw render. Prediction: the same darker, rounder B as the pulse pair, and B is no longer the quieter one. FALSIFIED if B at matched rms is no longer heard as darker/rounder (then loudness, not shape, carried the first pair), or if it is heard as louder or brighter.
|
||||
|
||||
Targets chosen from the calibration fit: white .33 -> HNR 20 dB; high-passed .33 -> 20 dB, .165 -> 25 dB, .6 -> 15 dB.
|
||||
|
||||
## Measured on main ca6af1d
|
||||
|
||||
Figures for the ten kit clips, taken on main `ca6af1d` (PR #52) with `hnr.py` and `clips.py` (`measures.txt`). Removed from the clip titles, which the listener can see through the export.
|
||||
|
||||
| clip id | file | HNR dB | output rms | peak | note |
|
||||
|---|---|---|---|---|---|
|
||||
| pulse-saw | saw_a.wav | 30.3 | 0.0537 | 0.227 | scaffold: band-limited saw through the tilt lowpass |
|
||||
| pulse-rosenberg | ros.wav | 29.5 | 0.0401 (-2.5 dB re saw) | 0.102 (-6.9 dB) | rms deficit is a fact at the output; the prediction row's "loudness the same" is already falsified there |
|
||||
| pulsematch-saw | saw_b.wav | 30.3 | 0.0537 | 0.227 | byte copy of pulse-saw's render |
|
||||
| pulsematch-rosenberg | ros_matched.wav | 29.5 | 0.0537 (matched to the saw) | 0.137 | ros times 1.341 (+2.55 dB) |
|
||||
| noise-clean | clean.wav | 30.3 | 0.0537 | 0.227 | the scaffold's noise floor, 0.05; noise 0 measured 30.8 before the tract fix, the method's ceiling here |
|
||||
| noise-hnr20 | w20.wav | 19.9 | 0.0539 | 0.242 | white noise raised to level 0.33 |
|
||||
| colour-white20 | w20_b.wav | 19.9 | 0.0539 | 0.242 | byte copy of noise-hnr20's render |
|
||||
| colour-hp20 | h20.wav | 19.9 | 0.0540 | 0.238 | noise high-passed at 2 kHz (2nd order), level 0.33 |
|
||||
| aspiration-hp25 | h25.wav | 25.8 | 0.0538 | 0.231 | high-passed noise, level 0.165 |
|
||||
| aspiration-hp15 | h15.wav | 14.9 | 0.0546 | 0.250 | high-passed noise, level 0.6 |
|
||||
|
||||
Band energy re total, 0-0.5 / 0.5-1.5 / 1.5-3 / 3-6 kHz, from the render before the tract fix of PR #52: saw -2.3 / -4.0 / -24.0 / -24.9 dB, Rosenberg -0.3 / -12.4 / -43.7 / -40.3 dB. These were not re-taken with the original method. A whole-file Hann periodogram on the new render gives saw -2.2 / -4.0 / -24.0 / -25.0 and Rosenberg -0.3 / -12.5 / -43.2 / -40.2, within 0.5 dB; a different FFT estimate of the new render differs by under 1 dB, and an independent check found the figures depend on the estimator (Welch differs by up to 1.6 dB, a raw periodogram by up to 3.1 dB). Read them as estimates to about 1 to 3 dB. The HNR figure is the Boersma autocorrelation method of `hnr.py` on the rendered output; Praat's autocorrelation HNR is within 0.5 dB for the noisy clips, and the clean clip sits at this method's ceiling.
|
||||
|
||||
Record corrections after the Soul pass (2026-11-14): the noise pair's table cell said the scaffold's noise was .04, the effective level is .05 (the cell was edited; the pair, the prediction and the falsifier are unchanged); the calibration addendum's count of renders was eight and is sixteen.
|
||||
|
|
|
|||
Loading…
Reference in a new issue