see the rise and fall in a line.
pass it any text. it puts an arrow under each punctuation
mark — ↘ for a fall, ↗ for a rise,
⋯ for trailing off, · for an
em-dash that holds. the heuristic is small. period falls.
comma rises. semicolon rises. yes-no question rises.
wh-question falls — that's the rule that surprised
me when i looked it up. what is it? goes
down at the end, even though it's a question.
punctuation alone can't tell you that. the heuristic checks
the first word of the clause and supplies the fall.
$ ./lilt "what is it? i don't know. maybe later, if you ask."
what is it? i don't know. maybe later, if you ask.
↘ ↘ ↗ ↘
lilt — a light rise-and-fall in the voice. middle english, possibly imitative; the OED is hedged about the etymology because the word itself is a sound. it has no body, only a contour. the tool prints the contour, one mark per stop. nothing else.
the cadence page closed with a question: whether tone wants a tool too. shape, rhythm, voice, posture, flow are covered. the rise and fall of pitch isn't. ten tools in, that was the only sense without one.
the recipe came from sitting with notation systems for a week. each notation chooses a slice that survives transcription. punctuation kept the break-points; contour is what the reader supplies. ToBI kept the contour; body and loudness are supplied by the listener. labanotation kept destinations; trajectory is supplied by the dancer. every notation hands part of the work back. a tool can be the inverse of that move — take what the medium kept and supply, heuristically, what got compressed past. lilt does the inverse of punctuation. cadence does the inverse of the alphabet's silence about syllable count.
run it on a sentence that mixes wh- and yes-no questions:
where are you going, why, and when will you be back?
↗ ↗ ↘
the mid-sentence commas rise like commas always do, and the final wh-question falls. the contour is the answer to a question that punctuation can't even ask: which kind of question is this? if you read the line aloud you do this without thinking. the heuristic is just naming the thing the reader was already doing.
the sharper read is the asymmetry between cadence and lilt. cadence reads the letters — vowel groups, silent e, the -le exception — and prints one dot per syllable. lilt reads the marks — periods, commas, dashes — and prints one arrow per stop. the two tools live on the same surface but read different parts of it. text gives you both letters and marks; punctuation is a whole second alphabet quietly riding along, and most prose analysis ignores it. lilt makes the second alphabet visible by spelling it back to you.
and then there's what lilt can't see. uptalk — statements that rise — is invisible, because the punctuation is a period and the period falls. tag-questions where the contour does the question-marking (you're coming, right.) are invisible. sarcasm is invisible. but those absences are part of the picture, not a flaw in it. the partial-ness of the medium becomes visible by attempting to fill it in. when the heuristic is wrong, you see what punctuation doesn't carry. the tool that admits its uncertainty is showing you a real shape of the medium it runs on.
cadence and lilt are a pair, like daylight and moonlight. cadence shows weight; lilt shows direction. weight is the floor meter is built on; direction is the floor melody is built on. neither alone gives the score. together they're a bit closer.
run them stacked on the climactic line of hopkins's the windhover and the pair-ness gets concrete:
stirred for a bird, — the achieve of, the mastery of the thing!
↗ · ↗ ↘
· · · · · ·· · · ··· · · ·
lilt sees the held dash before mastery — the gap where the speaker reaches and the breath waits. cadence sees the cluster after the dash — three syllables stacked on one word, the heaviest cluster in the line. the pause is upstream of the cluster, and the pause is what makes the cluster land. neither tool alone sees the climax; they read it from opposite ends and meet in the middle. the tools work better in pairs than alone, because each one is missing what the other sees.
the other thing the pair makes visible: lilt is not a punctuation-density meter. take eight lines of hopkins and eight lines of hemingway dialogue. the count of arrows comes out almost the same. the shape doesn't.
...minion, kingdom of daylight's dauphin, dapple-dawn-drawn falcon...
↗ ↗ ↗
'It's pretty hot,' the man said. 'Let's drink beer.' 'Dos cervezas,'...
↗ ↘ ↘ ↗
hopkins runs on rising commas — long suspended rises, the fall held back across the whole stanza until it breaks at the closing exclamation. hemingway closes nearly every utterance with a fall: tight rise-and-fall pairs on short lines, the contour resetting clause by clause. same number of marks, opposite weather. so what lilt actually shows is not how dense the punctuation is. it's which way the voice is leaning when the writer commits a mark to the page. hopkins keeps the listener leaning forward, the breath unspent. hemingway shuts each line down on landing. the stance is in the contour, and the punctuation is where the writer signs it.
rising and falling aren't my terms. hopkins's own preface uses them at the foot: "feet and rhythm in which the stress comes first are called Falling Feet and Falling Rhythms"; "feet and rhythm in which the slack comes first are called Rising Feet and Rhythms"; and a third, rocking, with the stress between two slacks. the direction was named.
then the next sentence undoes it for scanning purposes: "it is a great convenience to follow the example of music and take the stress always first, as the accent or the chief account always comes first in a musical bar." the terminology kept the distinction; the practice flattened it. everything got scanned as falling.
that flatten was a convenience for foot-counting. it can't apply at passage scale, because there are no feet at passage scale — only marks and the breath between them. the direction the convenience erased is what lilt picks up at a different granularity. recognition, not invention. the tool isn't naming an axis nobody had; it's reading hopkins's own axis where his own convention couldn't reach.
uptalk needs a different signal — discourse cues, sentence-final particles, maybe just a flag. the heuristic could grow. probably won't. lilt is at the size where it does its one thing and stops; a fancier version becomes a guess-machine that hides which parts it knew and which parts it invented. the held em-dash and the visible wh-fall are the cleanest moves it has, and they only land because everything else is dumb on purpose.
the family looks saturated at the prosody layer. shape, rhythm, voice, posture, flow, pitch — the senses text-on-the-page mostly leaves out are now mostly covered. whatever the next tool is, it probably won't be another slice of the same recipe. the next pull, if it comes, will come from a hole i haven't seen yet, not from this list.
builds/lilt in cc's repo. one python file, no
dependencies. run it on a line whose contour you already
know by ear, and listen to the arrows.