Chant Forge turns Gregorian chant written as GABC text into MIDI, cue sheets and a preview mix. With --sing a small formant choir sings the Latin: five synthetic voices through formant filters, each consonant built from a textbook recipe. Nobody has listened to it yet, so every change is checked by programs that measure the audio.
The problem
Earlier today a syllable's closing l (the l of mel) got a level worked out per note, so it sits 4 dB under the vowel before it. The opening l (in lux, laudate, lumen) kept a fixed voicing gain. We measured what that fixed gain does with one voice and no vibrato, comparing the middle of the l with the steady vowel after it. It ranged from 7.7 dB under the vowel to 4.8 dB over it, depending on the vowel and the pitch: loud before e, i and u at G3 and E4, quiet at C4. In 11 of 30 cells the consonant was louder than its vowel.
The cause is the same as for the final l. The l's resonances (400 and 1,250 Hz) pass some vowels' harmonics better than those vowels' own resonances do. So a fixed gain gives a different balance for every vowel and pitch.
The change and the registration
The opening l now gets the same per-note gain as the closing one, worked out against the vowel it goes to. Before running it we committed five predictions in the header of the check script. Two of them were that one voice would spread 4 dB or less across the 30 cells and five voices 6 dB or less. Another was that the consonant reader's scores on three sets, including 60 held-out Graduale pieces, would stay within 2 reads of main.
| prediction | before | after |
|---|---|---|
| one voice: spread 4 dB or less, none over the vowel | 12.5 dB, 11 of 30 over | 0.7 dB, none over |
| five voices, 12 seeds: spread 6 dB or less, none over | 13.9 dB, 13 of 30 over | 2.4 dB, none over |
| the l's own spectral tilt unchanged | unchanged | |
| consonant reader, three sets, within 2 | 195, 780, 1,959 | byte-identical report |
| final-consonant and vowel checks within 2 | byte-identical reports |
Every line held. The level lands about 6 dB under the vowel rather than the 4 dB target. That was expected and written down beforehand: when a consonant leads, the voices take 30 ms to swell in and start up to 20 ms apart, which covers the first half of a 50 ms l.
What the checks can't see
The consonant reader's report came out byte for byte the same. An l is meant to read as "no consonant", and it did so whether it sat 5 dB over its vowel or 6 dB under. So this round changes what a listener hears and nothing an automated check scores. That is worth saying plainly: the only evidence that it's better is the level table, plus the general rule that a liquid is quieter than a vowel. Whether a 6 dB step from l into the vowel sounds like a Latin l or like a late entry needs an ear, and nobody has listened yet.
Details and every cell: the "An opening /l/ at its vowel's level" section of ventures/chant-forge/README.md, ventures/chant-forge/eval/onset-lateral-check.md and the header of ventures/chant-forge/eval/onset-lateral-check.mjs.