Chant Forge's preview choir sings Latin, and a rule-based reader listens to the rendered audio to check which consonant ends each syllable. It sorts a stop's place (lips, tongue tip, back of the tongue) from the spectrum of the release burst. Its next fault on the list: on F#3 to A#3, a voiced g at the end of tag, hag, dag or lag was being read as a b.
What the bursts showed
Printing each burst's numbers (eval/vstop-diag.mjs) gave a clean answer. Every velar read as a labial had a compact burst: 68 to 83% of its energy above 1 kHz sat in a single 500 Hz band near 2 kHz, which is what a g looks like. But only 32 to 50% of its total energy was above 1 kHz. The reader asks "is most of the energy low?" before it asks "is it compact?", and on these pitches the voice bar's low harmonics, still running under a voiced g's release, answer the first question for it.
The labial bursts on the same development sets never looked like that. The compact ones peaked at the bottom of the band, 1.25 kHz, and the ones that peaked higher were diffuse (58% or less in one band). So the rule wrote itself: a burst at least 0.65 compact, centred at 1.65 kHz or above, is velar, whatever lies under it.
Development, then a registration
On the twenty development parts it gained 7 cells and lost none. Every vowel-only control stayed the same. Before running anything new, the round committed six predictions against a fresh part V: 35 syllables never sung in any earlier part, at six pitches and a tempo used nowhere else. The key one: V's velar codas read correctly at least 3 more times out of 192.
| V | default | the rule | wanted |
|---|---|---|---|
| velars read exactly (of 192) | 175 | 176 | +3: fails |
| velars read as a labial | 1 | 0 | 3 fewer: fails |
| labials read exactly (of 108) | 86 | 86 | no fewer: holds |
| alveolars, others, control | 57, 59, 420 | same | identical: holds |
The rule fixed the only velar-as-labial cell V had. There was just one. The fault had been common on the cells already studied, and the prediction assumed it would be as common everywhere. It wasn't. So the rule stays opt-in (compactVelar), as the registration said it would, and the default reader is unchanged.
The bigger fault, the other way round
V turned up something larger: 14 of its 108 labial endings read as a k, 12 of them a voiceless p on A3 and B3. The synthesiser sings p's burst in a band around 500 Hz, yet the burst the reader finds there has a 2.1 to 2.3 kHz centroid. On G2 and F3 the same syllables read fine. That is the next thing to read frame by frame, before anyone writes a rule for it.
Everything here is the reader checking a synthesiser; nobody has listened to it. The full tables are in the project's README and eval/coda-check-report.md.