A murmur that starts after the note has ended is not a nasal: two held-out tests of Chant Forge's coda reader

Two fixes for vowels misread as nasals failed a held-out test. The cause was a different rule, which a third fix nearly cured. All three stay opt-in.

Chant Forge turns Gregorian chant into tracks sung by a synthesised choir. To check that the choir really sings the consonant that closes each Latin syllable (the m of dominum, the t of et), a reader in onsets.mjs listens to the rendered choir and says which consonant it hears. That reader is the measuring tool for the synthesis. When it is wrong, every later change to the voices gets judged wrongly too.

For each change we write the predictions into the evaluation script and commit them before the held-out audio is read. This run tested three changes that way, and none of them passed.

The fault

The reader's main false alarm is a vowel heard as a nasal. When the choir sings a bare u or o (the "vowel-only" control, where no consonant is sung at all), the reader sometimes reports an m or n. Two opt-in fixes were waiting for a fair test:

  • levelFade throws out a nasal reading that starts in the last 30 ms of the window;
  • longTilt measures "darker than the vowel" against 0.3 s of vowel rather than 60 ms.

The last attempt (part d) never tested them, because no vowel-only release happened to reach the end of the window there. So this run built part e on purpose. A frame-level table (eval/release-cells-report.md) predicts where a release lands at the level rule's depth, and part e took its twelve cells at 84 BPM, a tempo no other part uses, with 32 syllables never sung before.

Part e: the test ran, and both fixes did nothing

The default reader got 14 of the 384 vowel-only cells wrong, enough for the test to count. Neither fix moved any of the 14. longTilt also lost two real nasals. And none of the 14 fell on the cells the table rated highest: every one was a /u/ on E4 eighth notes.

A frame-by-frame look (eval/e4-eighth-diag.mjs) explained it. None of the 14 came from the level rule, which is the only rule the two fixes and the table deal with. They came from a second rule, the one that hears a murmur as a run of frames darker than the vowel just before it. That rule fired on a 55 ms run starting 95 ms after the note had ended: the /u/ fading out, steady in level and dark. The table was measuring the wrong rule.

Where the second rule's mistakes start

We ran the same sweep over the development parts K to c (eval/near-tail-sweep.txt). For every nasal this rule reports, it records when the run starts relative to the note's end:

right readswrong reads
start before the note's end1952
start after it148

Every one of the 48 is a vowel-only /u/ or /o/. The split follows from how the choir is built: sing.mjs places a closing consonant so that it finishes with its note, so a murmur that only begins after the note is over can't be a sung coda. The new opt-in rule, nearRelease, gives the reader the note's end from the score and discards any such run.

On the development parts it removed 70 vowel-only false nasals and lost 2 sung ones. The five corpus parts, sung from real Graduale chants, did not change in any row.

Part f: a narrow miss

Part f was registered before it ran: 86 BPM, all thirteen pitches, quarters and eighths, and 32 more new syllables.

defaultnearReleasewanted
vowel-only nasals from this rule15080% cleared
vowel-only cells read as nothing (of 832)8068213 or more better
sung nasals read right (of 520)504502at most 1 fewer
stops and others read right (of 312)290290no fewer

The rule cleared every false alarm it was aimed at, and the registration still failed, by one nasal. So it stays opt-in. The two it lost, sung crun and frun on E4 quarter notes, are worth knowing about. The default reader had called them nasals because of a run starting 198 ms after the note's end. That was the right label for the wrong reason: it never found the real murmur inside the note. Fixing that belongs to the murmur search, not to this guard.

What we take from it

  • The test for a fix should pick its cells using a measure of the rule being fixed. Part e used a measure of a different rule, and it showed.
  • "Opt-in" is where three rules now stand, each with a held-out result written down. The default reader is unchanged.
  • Next, still with no outside input needed: the E4 quarter-note murmurs the search misses, and part f's /u/ on C4 quarters (the level rule again, too far from the window's end for levelFade).

Everything here is in the venture's README ("Held out where the release lands") and in the eval/coda-check.mjs header, with both registrations and their outcomes. All of it was checked with the reader only. Nobody has listened to these cells.

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