Sunt read as sut: the n a synthetic chant choir sang and its own checker missed

Chant Forge's choir sang the n in sunt, but its acoustic checker heard only the t. One cue fixed it: a murmur starts with a step, not a ramp.

Chant Forge turns Gregorian chant into tracks for the gym, with a synthesised men's choir singing the Latin. Nobody has listened to that choir yet, so every change to it is checked by machine. A small acoustic reader (onsets.mjs) listens to the rendered audio and writes down which consonant it hears at the end of each syllable. The reader never sees the synthesiser's settings.

Before today it kept writing sunt down as sut.

The n was there

Read frame by frame, the n in sunt, ant and mens is plainly in the audio. It is a nasal murmur of 50 to 70 ms that the t's closure cuts off. Through it, the energy above 1 kHz drops 17 to 55 dB against the band under 500 Hz, compared with the vowel just before. The reader's three nasal rules each missed it for a different reason:

  • Too loud. The level rule wants a murmur at least 5 dB quieter than its vowel. The choir has five slightly detuned voices whose low harmonics beat against each other, and on low notes that leaves the murmur only 2 to 4 dB under the vowel.
  • No step under 500 Hz. That rule was written for -um after a dark vowel, and the n of -unt doesn't drop in that band.
  • Not steady enough. The newest rule wants the murmur's darkness to hold within 4 dB, and here it wobbles by about 6.

So the reader skipped the n and found the next thing it recognised, the t.

Step, not ramp

The obvious fix ("anything dark just before a closure is a nasal") was the first draft, and it failed on a synthetic set of 39 syllables at six pitches. It read up, uc, ub and ap as nasals 14 times. A vowel darkens on the way into a closure too. That happens most of all to /u/, whose resonances slide toward the lips or the soft palate.

The frame-by-frame numbers showed the difference. A vowel darkens as a ramp, 2 to 5 dB per 5 ms hop. A murmur darkens as a step: the nose takes over in one hop, 8 to 15 dB at once, and then holds. So the new rule asks for a dark run of 40 ms or more that starts with a step of 8 dB within two hops and lasts right up to the obstruent. On the synthetic set, cluster nasals went from 46 to 124 of 132, and not one other syllable changed its reading.

Written down before it was run

The thresholds were chosen on synthetic syllables only. Every corpus set had already been used in earlier rounds, so none of them was clean held-out data any more. Instead, five predictions went into the check script's header and were committed before any corpus set was run with the rule. All five held:

beforeafter
unt/ant in the tuning set7 of 1212 of 12
nasals in 51 scan-checked pieces106 of 137116 of 137
corpus cluster nasals (five sets)7 of 1111 of 11
vowel-only control, every setidentical

What it doesn't tell you

Nothing in the synthesis changed. The checker can now see an n that was already being sung. Whether a listener hears sunt or sut, and whether five real voices beat the way five synthetic ones do, still needs a person with headphones. The details, including the syllables still missed (nunc and mens on high notes, where the murmur arrives without a step), are in the project's README under "-unt read on its stop". The code is MIT licensed.

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