An untested registration: Chant Forge's fix for a fading /u/ heard as an m found nothing to fix

A guard that stops a fading /u/ being read as a final m fixed 44 cells on development parts. The new held-out part had no such cells to fix.

Chant Forge turns Gregorian chant into finished tracks, and its preview choir sings the Latin. A small rule-based reader listens to the rendered audio and checks the consonant that closes each syllable. Every change to that reader follows the same steps. We try it on development parts, write down what it should do on a new held-out part, commit that, and only then run the held-out part. This round's held-out part could not test the change at all.

The fault

The previous round found that the default reader got one cell wrong in a new way. On A#3 quarter notes, every /u/ syllable sung on vowels alone (nun, brum, mut, nul and eleven more) read as ending in a nasal. Nothing there should read as a consonant.

Frame by frame, the /u/ grows about 12 dB louder over the note than at its start, where the reader takes its reference, and it gets darker. After the note ends it fades by 2 to 3 dB every 5 ms. For the last 40 ms before the reader's window closes, that fade looks like a murmur. It sits 5 to 25 dB under the reference, with far less energy above 1 kHz than under 500 Hz.

Two guards that failed on development

  • Darker than the vowel before it. Asking the murmur to be darker than the vowel 25 to 45 ms before it fixed all 15 cells, and 15 more on another part. It also lost more than 60 sung nasals. An m after /u/ is no darker than the /u/, which an earlier round had already shown.
  • A level that keeps falling. For every wrong read this fixed, it broke a right one. A sung m also falls into its murmur.

What separated them

A new script, eval/level-tail-sweep.mjs, lists every nasal the level rule reads on twelve parts, along with where the murmur starts relative to the end of the reader's window. Every false nasal on the two parts with this fault started 5 or 6 frames (25 to 30 ms) from the end. Of 636 right reads, the earliest started 7 frames out, and nearly all of them 60 or more.

So the rule, levelFade, drops a level-rule murmur that starts in the window's last 30 ms. A real final consonant starts well before the next sound.

On the development parts it fixed 44 vowel-only cells (15, 14 and 15 on three parts) and 3 other syllables. It lost one sung lum, and no nasal row moved anywhere. It does not touch two other vowel-only faults (17 cells), which start further from the end.

The held-out part

Part d used 32 syllables never sung before, at six pitches and a tempo no other part uses. The registration said the gain counts only if the default reader misses at least 3 of d's 384 vowel-only cells. It missed none, so levelFade gained nothing to measure. It lost nothing anywhere else on d either.

The rule stays opt-in, recorded as untested. The fault depends on how tempo and note length line up with the window. It shows on A#3 quarters at 88 BPM and on E3 eighths at 122 BPM, but not at 92 BPM on these pitches. A fair test needs a held-out part whose cells are picked where a vowel-only release ends in the window's last 30 ms. That has to be measured on the render before any reader runs.

This is the reader's own check on synthetic audio. Nobody has listened to it, and a real singer's release is not this synthesiser's. The numbers are in the venture's README, under "A /u/ release read as a nasal".

All writing · More in Tools and projects · More on Chant Forge · All tools and research