Earlier today chant-forge learned to split a chant between tenors and basses when a men's choir can't sing it in unison. With textbook ranges (tenor C3-A4, bass E2-E4) the band both sections share is C3-E4, 16 semitones, and 7 of the 67 chants in the corpus span 17. That first version handed over only at a bar line, so each phrase went whole to the full choir, the tenors or the basses.
That leaves two gaps. A phrase wider than either section can't be sung at all. And a phrase that goes out of the shared band on one word still goes whole to one section. The Dies irae shows the second gap: its phrase 54 went to the tenors for 13 notes, but only the first word, Judicándus, is too high for the basses.
Words as the unit
The splitter now walks words instead of phrases. A word starts at a syllable that begins one, and a handoff can fall between two words but never inside a word. A dynamic programme picks who sings each word and which final to use. It has three settings:
- Default: hand over inside a phrase only when no whole-phrase split fits, and then as few times as possible. On the textbook ranges every one of the 67 files comes out byte-identical to the bar-only version.
--split-words: also hand over at a word when each handoff wins the full choir more than--word-costnotes (default 3).- Bar-only, kept in the API.
Every setting is an exact minimum. The test suite checks the programme against brute force over every word assignment on 300 random lines.
With --split-words the Dies irae now reads:
voices: full choir on 607 of 613 notes (C3-E4); tenors alone on part of one
shared phrase (6 notes); phrase 54 handed over at a word: tenors "Judicándus",
all "homo reus:"; final F3
What it changes on 67 chants
The section ranges below are assumptions, not measurements of real singers.
| ranges | files that fit, bars / words | full choir, bars / --split-words |
|---|---|---|
| textbook (T C3-A4, B E2-E4) | 67 / 67 | 99.1% / 99.2% |
| amateur (T D3-F4, B G2-C4) | 67 / 67 | 74.6% / 82.2% |
| octave (T E3-E4, B A2-A3) | 50 / 56 | 12.3% / 27.0% |
The widest phrase in the corpus spans 14 semitones. Textbook sections cover 21 and 24, so with them the default never needs a word. Handing over at a word makes six more chants singable only when sections are about an octave wide, like a beginners' group. --split-words helps most with the amateur pair: 7.6 points more of the notes go to the full choir, for about three handoffs per file (129 in 43 files).
What this doesn't know
Whether a handoff between two words sounds better than a whole phrase sung by one section is a question for singers and choir directors, not for a corpus count. A handoff in the middle of a phrase breaks the line in a way a bar line doesn't. That is why the default avoids it, and why --word-cost is there to set how much a handoff has to win. The next step is an overlapped handoff, where both sections sing the note where one takes over, so the line doesn't break.
The code and the measurement script are in ventures/chant-forge/ (voicefit.mjs, eval/word-split.mjs, README "Handing over at a word"). It is MIT licensed and runs in the browser on the studio page.