Chant Forge turns Gregorian chant written as GABC text into MIDI, scores and a preview mix. The preview has a small formant-synthesis choir (sing.mjs, no dependencies) that sings the Latin: vowels, opening consonants and, since this week, the consonants that close each syllable. An acoustic reader (onsets.mjs) listens to the rendered audio and checks what was sung.
The fault
An earlier run found that the sung /l/ was louder than the vowel before it. A fixed voicing gain of 1.4 put the /l/'s steady part between 1 dB under and 8 dB over its vowel. It was loudest after e, i and u, because the /l/'s first two resonances (400 Hz and 1250 Hz) pass those vowels' harmonics better than the vowels' own resonances do. Real /l/s are weaker than the vowels around them, so every -el or -ul would have come out as a small swell.
The change
Final nasals already had a per-note gain. It sums the power of the source's harmonics through the vowel's resonators and through the murmur's, then sets the murmur 7 dB under. The /l/ now gets the same treatment, with a target 4 dB under its vowel. That 4 dB is an assumption, picked to sit between a vowel and a nasal. Nobody fitted it to recordings, and nobody has listened to the result yet. With one voice and no vibrato the /l/ now measures 3.8 to 4.0 dB under its vowel, at four pitches and after all five vowels.
The prediction that failed
Before the first run, the check script's header wrote down four predictions. The first said that, measured on one seed, every /l/ would sit 1 to 7 dB under its vowel. That held on 20 of 30 notes, up from 9 of 30 before the change. As registered, it fails.
The failure came from the measurement, not the gain. The /l/'s steady middle lasts only 25 ms. Over a span that short, the choir's five slightly detuned voices beat against each other, so one seed's reading swings about 5 dB either way from note to note. With the power summed over twelve seeds, the old gain put the /l/ anywhere from 5.4 dB under to 8.6 dB over its vowel. The new gain puts it 2.4 to 5.0 dB under, with a mean of 3.7. We tried a gain that also averaged over the detuned voices and the vibrato. It moved nothing by more than 0.2 dB, so it was dropped.
The other predictions:
- The /l/'s spectral shape is unchanged.
- The final-consonant reader's results are identical on every set, including three held-out sets.
- The opening reader lost one read of 934: in the Dies irae, a syllable opening on its vowel now reads as a nasal.
The full tables are in ventures/chant-forge/eval/lateral-level-check.md.
Chant Forge is MIT licensed. Its preview choir is a sketch for arranging, not a recording, and nobody has yet heard this change.