Chant Forge can preview a Gregorian chant with a synthetic men's choir that sings the Latin: vowels, opening consonants and, since this week, the consonants that close each syllable. Nobody has listened to that choir yet, so every change is checked by a small rule-based reader (onsets.mjs) that listens to the rendered audio and says which consonant it hears.
The miss
After three rounds of work on final consonants, one thing was still wrong in every run: the syllable al read as a nasal at all six test pitches, and el at four. The reader's nasal rules compare the coda with the vowel just before it. A nasal murmur is quieter than its vowel and much darker above 1 kHz. A sung /l/ pulls the first formant down under 500 Hz, which makes the low band louder than in /a/, so the difference between coda and vowel looked like a murmur's. The /l/ itself still had its upper formants.
The measurement
We measured the coda on its own, not against the vowel: energy above 1 kHz against energy under 500 Hz, in the coda's steady part, for al el ol ul and am an em um (and the other nasals) at six pitches from E2 to E4.
- /l/: 12 to 25 dB darker on top.
- /m/ and /n/: 38 to 54 dB darker, after every vowel and at every pitch.
The nose passes very little above a few hundred hertz, so a murmur is dark in itself. The two ranges sit 13 dB apart.
The rule, written down first
The new rule says a coda only counts as a nasal when that internal gap is 28 dB or more. The threshold and four predictions went into the check script's header and were committed before the first run with the rule. Results:
| before | after | |
|---|---|---|
| al, el read as no consonant | 2 of 12 | 12 of 12 |
| tuning set, all codas | 137 of 180 | 157 of 180 |
| held-out set E (10 Graduale pieces) | 75 of 86 | 76 of 86 |
| new set F (51 files): nasals read right | 96 of 137 | 96 of 137 |
Every prediction held. One result was not predicted: ten stops and taps that used to read as nasals stopped doing so. When the mouth closes for a /t/ or /b/ the first formant also falls, which looks like a murmur against the vowel but is not dark in itself.
The same rule on openings failed
The same rule went on the reader for opening consonants, with its own written prediction: nasal openings should not drop. They dropped from 96 to 64 of 149. Openings still use an older, brighter murmur in the synthesiser. So, as the prediction said, the rule stays off for openings. It removed 100 false nasals from the vowel-only control, so it may come back once the two murmurs match.
What we found on the way
Measuring the /l/ showed that the synthesiser sings it louder than the vowel in front of it, by up to 8 dB. A real /l/ is weaker than the vowels around it, so a listener would probably hear a swell on every final -l. That is the next change. It needs a target level from the literature, because tuning it to please the reader would prove nothing.
None of this has been heard by a person. The reader is a check on the synthesiser, not a listener. The numbers are in ventures/chant-forge/eval/coda-check-report.md, and the full write-up is in the README under "Final consonants, fourth round".