A chant-reader fix that worked on its training set and failed two held-out tests

A chant consonant reader missed nasals on high short notes. The fix cleared every training miss, failed two pre-registered held-out sets, and ships off.

Chant Forge renders Gregorian chant with a synthetic choir that sings the Latin, and a small rule-based reader listens back to check that the final consonants are really in the audio. On the set the reader was tuned on, it got every cluster syllable right except four, all on the same kind of note: an E4 eighth note. Nunc, tunc and mens read as a voiced b, and sunt as a t. The n was missing.

Read in 5 ms frames, the n was there after all. It is just very quiet. On a 330 Hz voice only one harmonic falls under 500 Hz, so the murmur sits 19 to 32 dB under the vowel, deeper than any nasal rule allowed. Inside it, the five slightly detuned voices beat: they drift in and out of phase, and at one dip the murmur fell to -31 dB. To the reader that dip looked like a closure, and the swell back up looked like a stop's burst.

The rule

The fix reads what a murmur does that nothing else does. It holds at -12 to -35 dB for 60 ms or more, then ends in an obstruent: either noise that outlasts it (the s of mens), or a closure that goes properly silent and then bursts (the k of nunc). A voiced stop's voice bar never goes silent before its burst, and a vowel falling into a t passes through that level band in 15 to 35 ms. On the training set the rule fixed all four misses and broke nothing.

Two held-out tests

The project's habit, kept here, is to write a fresh test set and commit its predictions before the rule is run on it. The first set, part N, used new syllables at new pitches and 108 BPM. The rule failed three of its five predictions. N's high eighth notes were already read right 19 times out of 20 without the rule, because an eighth at 108 BPM lasts 0.28 s rather than 0.25. Voiced stops were read as nasals 5 times instead of once. Three other parts each lost a syllable.

The frames showed two faults. On a short note the reference vowel is measured 45 ms in, which is still the attack, so a voice bar looked only 24 dB down. And after the fading tail of a dark /u/, a voice bar counted as "18 dB under the run", the rule's first definition of silence. The second version compares against the vowel's own peak just before the run and asks for real silence: a frame at -50 dB. It went back over the training set and N without a single loss, and a second fresh set, part O, was written and committed: 126 BPM, eighths of 0.24 s, and one pitch no part had used.

It failed again, on three predictions:

predictionwithout the rulewith it
O's high eighth-note clusters, at least 2 more25 of 3025 of 30
O's voiced stops read as a nasal, no more34
every earlier part no lowerone part 106 to 105

Added up over every part, the rule gains eight reads and loses one. That sounds like a win. But it was registered as a fix for short high notes, and on the one set built to test that claim it fixed none of them. Its gains came from other syllables. A net gain found after two rounds of reading the data is exactly what tuning on the test set would also produce. So it ships the way it was registered: off by default, one switch away (deepMurmur: true, or --deep on for the whole report), with both columns printed for N and O.

What is next

O's five high eighth-note misses (punc read as a k, tunt and ment as a t, sons as an s) have not been read frame by frame yet. They need their own explanation and, because their misses are now printed, a part P that nobody has seen. All of this is the reader checking the synthesis. Nobody has listened to whether a choir on E4 sings nunc or nuc.

The measurements are in the Chant Forge README ("E4 eighth notes: a murmur read as a closure") and the coda report.

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