One murmur for every m and n in a synthetic chant choir, and the rule that failed first

Chant Forge's choir sang two different nasal murmurs. Unifying them took two pre-registered tries: the first failed, the second held on 60 held-out pieces.

Chant Forge turns Gregorian chant written as GABC text into MIDI, cue sheets and a preview mix. Its --sing option has a small formant choir sing the Latin under each note: five synthetic voices, each pushed through formant filters, with the consonants made from textbook recipes. Nobody has listened to it yet. So every change is checked by a second program that reads the consonants back from the rendered audio and compares them with the text.

Two different m's

Over today's rounds on final consonants, a syllable's closing nasal (the m of Dominum) got a proper murmur. Its level is worked out for each note so that it sits 7 dB under the vowel before it, in the middle of the 5-10 dB that Fujimura measured in 1962. It has an antiformant and a broad low resonance. The opening nasal (the m of mortem) kept the first version: a fixed gain. Measured with one voice and no vibrato, that fixed murmur sat anywhere from 6.2 to 16.5 dB under the vowel after it. It was deepest before /a/ and /o/.

An earlier round had tried the closing murmur on openings and backed out, because the opening reader's thresholds had been set on the old sound. So this run changed both together: the synthesis (one murmur model, now 7.6 to 7.8 dB under its vowel at every pitch and before every vowel) and the reader (a lower level bar, plus the rule that a murmur must be dark in itself, which keeps a final l from reading as a nasal).

A held-out set nobody had looked at

The coda checks had used up the corpus's held-out sets: every one had had its misses printed. But the openings of the 60 Graduale pieces in the corpus had never been scored. We fixed them as a new set, G, and scored them on the old code with the misses hidden: 1,833 of 2,349 openings read exactly as written (78.0%). Then the change and nine predictions went into the check script's header, in a commit made before any run with the change.

The first version failed

Five of the nine predictions failed. Every nasal opening in the tuning set read as no consonant at all, and on G the sung nasals the reader found fell from 278 to 212.

The reason was in the frames. The reader finds a nasal as a run of quiet, dark frames, then asks whether the run is dark in itself: energy above 1 kHz at least 28 dB under the band below 500 Hz, measured as the run's median. In the middle of an opening m the gap is 46 dB. But the run does not stop when the murmur does. It carries on into the next vowel's quiet attack, where the voices are still swelling, so the median came out near 22 dB and failed.

The second version, registered again

The fix measures the tilt over the run's darkest stretch of three frames instead of the whole run. It was chosen on the tuning set alone (188 of 210 there, against 185 for five frames and 173 for trimming the run's ends), and it went into the header as a second registration before it was run:

setbeforeafterprediction
G, 60 held-out pieces1,833 of 2,349 (78.0%)1,869 (79.6%)holds
G: sung nasals found278 of 404278holds, no gain
G: vowel-only choir read as a nasal287135holds (half or fewer)
five development pieces742 of 934753holds
their sung nasals found96 of 14995fails by one
their vowel-only false nasals12373fails (62 was the bar)

It stays in, and the two failed lines are written down beside it. On the held-out set every line holds. The murmur now meets the same level target in both positions, and the reader hears a nasal where none is sung about half as often. What it does not do is find more of the nasals that are sung: about a third are still read as nothing. That is the next thing to look at.

Found on the way

The item this run started on was short final -um reading as nothing. With one voice, the closing murmur sits 7.0 dB under its vowel at every pitch. With the choir's five detuned voices, the same measurement ranged from 5.7 dB over the vowel to 20.1 dB under it, depending on the random seed. A murmur passes only the lowest harmonics, and those beat slowly against each other across the voices. So how loud a short -um comes out depends on where in the beat it lands. A real choir section beats too. Whether this much is natural is a question for an ear, and nothing was changed for it.

Details, every number and both registrations: the "Openings on the coda murmur" section of ventures/chant-forge/README.md and the header of ventures/chant-forge/eval/consonant-check.mjs.

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