A slow lip closure, and a held-out part that missed for a different reason

A rule for slowly closed m's gained 16 cells on development parts and 2 on its held-out part. The held-out misses turned out to be a different shape.

Chant Forge turns Gregorian chant into finished tracks, and its preview choir sings the Latin. A small rule-based reader checks the sung consonants from the rendered audio, the way a listener would. A rule that changes the reader is tried on development parts first. Then it is judged on a new held-out part, with its predictions committed to git before that part is ever run. This is the twenty-fifth round of that loop. Its rule failed, and the way it failed was useful.

What the last round left

The round before left 11 cells of part b where a final m or n after /u/ was sung but read as nothing: num three times, mum, bun and cun twice each, run and vun once, mostly on D#4 and E4. The default rule for these, lipShut, looks for the moment the lips meet. The band above 1 kHz drops 12 dB or more within two 5 ms hops, and then the level holds within 4 dB for 30 ms.

Read frame by frame, num on an E4 quarter note does both, just more slowly. Its upper band falls 0, -3, -5, -9, -13, -15 dB against the vowel, which takes five hops. Its level then slides from -2 to -8 dB before it settles. The 30 ms hold, measured from the fall, misses by 2 dB.

The rule, and the guard it needed

The slow form counts the fall from any of the four frames before it, and lets the hold start up to two frames late. The first version read 6 of part A's vowel-only /u/ cells on C4 as nasals, and 15 of part J's controls. A /u/ fading at a note's end dips 15 dB in that band too, and then its vibrato brings it back 8 dB. So a late hold now counts only when the band never climbs back more than 4 dB over its lowest point since the fall. With that guard, every vowel-only control on every development part read the same as before.

Development, A to b: about 16 cells gained (b alone 133 to 137 of 144) and two lost, both a stop or bare vowel read as a nasal.

Held out: part c

Part c was written before it ran. It has 12 nasals after /u/ or /o/, chosen to look like b's misses (nasal, labial and cluster onsets), plus stops, bright vowels, clusters and controls, none sung in any earlier part, at 88 BPM.

cdefaultslow formwanted
nasals after /u/ or /o/ (of 144)1231253 or more better: fails
bright-vowel nasals, n-clusters (of 48 each)48, 4448, 45no fewer: holds
voiced, voiceless stops, others60, 36, 4860, 36, 48no fewer: holds
cells read as a false nasal1515at most 1 more: holds
vowel-only control (of 384)369369identical: holds

One short of the bar, so the slow form ships opt-in and the default reader is unchanged.

Why it missed

The default missed 21 of c's dark-vowel nasals, plenty to test the rule on. But 18 of those 21 were not slow closures. They were two-consonant onsets before -um: brum 6, glum 5, stum 4, flum 3. The slow form takes none of them back. It was written for one shape, and the held-out part's misses were mostly another. The last two failed rounds made the same mistake: they took the shape of the development misses for the shape of misses everywhere. Moving the reference vowel 100 ms later fixes only 2 of these, so a reference landing in the onset is not the cause. These cluster onsets are the next thing to read frame by frame.

Part c also caught a fault in the default reader that 26 earlier parts did not. All 15 of its vowel-only misses are in one cell, A#3 quarter notes. There, every /u/ syllable sung on its vowel alone reads as a nasal by the plain level rule. The bright vowels and /o/ in the same cell read correctly.

Everything is in the repository. The frames come from node eval/dark-nasal-diag.mjs b num and ... c brum. The registration and its outcome are in the header of eval/coda-check.mjs. The numbers are in eval/coda-check-report.md and the README, "A dark-vowel m that closes slowly". This round changed only the reader. The synthesis did not change, and nobody has listened to these cells.

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