The Tear Has Its Reasons: a Latin aria, a Neapolitan conversation, and feeling drawn as geometry

Watch on YouTube. The Tear Has Its Reasons. Lacrima Rationes Habet. A song about what a feeling looks like once you can see it.

Licensed CC BY 4.0.

Pascal wrote that the heart has its reasons, which reason knows nothing of. Blake wrote that a tear is an intellectual thing. Both lines were already in the song. Put them together and you get the title. The tear has its reasons.

A singer who never existed

The song needed a woman's voice. Soft. An ancient language. Something like a Bond aria, but kind.

ElevenLabs cannot sing. So every Latin word was spoken by three different voices, then re-sung through the WORLD vocoder onto the score. The vowels stretch. The pitch follows the melody, with a slow vibrato that only arrives on long notes. The consonants stay human.

Then the three became one. She is a wavetable singer: one pitch, one breath, and a throat that moves. Soft at the start. Velvet in the middle. Dark at her highest note, an F sharp at 2:38. Then soft again for the goodbye. Five Latin words carry the whole part. Lacrima. Rationes. Habet. Cor. And a hum.

Two men talking in Neapolitan

Under the music, OG Bobby and Roberto have a conversation. Fourteen short lines, low and easy, like two friends at a table while the band plays.

Bobby asks if he hears the sound. Roberto says yes, but not with his ears. With what, then? Wait. First it is air that moves. Then it becomes music. Then it goes inside you and never leaves. Like a tear. What cannot be said, you dance. And what cannot be danced? You sing it.

The last line is the title, in Neapolitan. 'A lacrema tene 'e raggiune soje.

Every line starts on a beat. The levels sit within four decibels of each other, so nobody shouts and nobody disappears.

Words that belong

Bobby speaks single English words through the piece. Heart. Tear. Mind. Body. Love. Hate. Respect. Dance it. The first mix had them riding on top. The final pass in ChipForge measured every word and brought each one to the same family as the aria and the Italian: three to four decibels over the music, a soft onset, and a tail that decays into the same room. Now they sound like they were always part of it.

Feeling, drawn as geometry

Every motion in the picture is read from the audio. Nothing is keyed by hand.

A flower of life grows one circle on each downbeat. Lissajous figures draw the ratio of each chord: 3 to 2 for D, 6 to 5 for the minors, 11 to 8 for the Lydian colour. The circle of fifths turns as a tilted orbit, with her note lit as she sings it. Her voice is a ribbon of light coloured by her throat.

Bobby's words are real curves, drawn on as he says them. The heart is a cardioid. The tear is a teardrop. Mind and body are two circles that join into a vesica piscis. Bobby's lines send amber waves from the left, Roberto's send blue waves from the right, and where they cross, they interfere.

At the climax the tear becomes a lens and splits the light into a spectrum. In the goodbye it falls and touches the mirror.

The finished film is checked against the sound. Every word and every Italian line lands within half a frame.

The house

The floor under the stage looked like a stage. So it got an audience.

Velvet curtains part on Pascal's line and close on the last note. Between the stage and you sit curved rows of seats in perspective, and every seat is a number. A live one. The machine is computing them from the music as it plays: how fast the loudness is rising, her frequency in hertz, how fast her pitch is moving, the phase of the beat, the ratio of the chord. Information at the speed of a machine thinking.

Each seat hears the stage a little late, by its distance. So every word breaks on the back row exactly on the sound and rolls toward you. In the dance the house bobs in a wave. In the breath it freezes and goes dark. At her highest note every seat shows the same number, 739.99, her F sharp. One crowd, one note. When the tear touches the mirror, the ripple runs through the seats and each one settles to 0.000.

The rate of change goes to zero. Stillness. Then the curtain.

The face

The thumbnail is OG Bobby as Anubis, the jackal, with human eyes. One amber, one blue, the two colours of the conversation. Bobby and Roberto are reflected in each eye. One gold tear, with the flower of life inside it.

It is the whole film in one frame. And it is the first frame of the video.

How it was steered

One afternoon into one evening. Two AI sessions on one laptop, one per repo. ChipForge built the music. Napkin Films built the voices, the picture and the release. I moved between them, listening, watching and deciding.

Both transcripts were kept. This is what they say.

Wall clock	7.6 hours, one day
My prompts	93 across both sessions
My words	about 2,300, a typical prompt 12 to 15 words
Prompts sent while the machine was still working	59 of 93
Words the agents wrote back	about 15,700
Tool calls	436
Code written	about 5,900 lines: 3,700 for the music, 2,300 for the voices and picture
Audio renders	over 100, plus 62 voice takes
Auditions I chose between	21
Thumbnail images generated	16
Messages between the two sessions	5
Time between turns: listening, watching, deciding	about 4 hours

And what kind of steering each prompt was:

Vision and direction	26
Taste and correction	28
Orchestration: handoffs, keys, status	18
Choosing between options	13
Audience and reach	7
Rights	1

Feel in, numbers out

Almost none of the direction was a specification. It was about feel: gentler, more human, more nuanced, balanced, part of the song, in time with the music. The machine's real job was translation. Every word of feel had to become a number it could measure and then hit.

A spoken word that should feel like part of the song became a target: three to four decibels over the music, a soft onset, and a tail that decays into the same room instead of stopping dead. Balanced became fourteen Italian lines within four decibels of each other. In time became within half a frame, checked on the finished film. Once a feeling had a number, the note never had to be given twice.

The timecode is the sharpest tool

Only 8 of the 93 prompts pointed at a moment in the song. Fifteen timecodes in all. Every one of them became a fix: a stepped-down note at 1:14, a clip cut off at 1:16, a phrase that caught one word too many at 1:23, a muddy word at 1:38. A listener pointing at a second does more than any general note.

The best ideas came late

The aria was not in the plan when the music started. It arrived as a passing thought in the third hour. The Neapolitan conversation arrived while the final mix was already running in the other session. The audience of numbers arrived after the film was finished.

Each one restructured the work. That is only affordable because nothing is baked: the music is stems, the picture reads a features file, and every gate can be rerun. A late idea costs a rerun, not a rebuild. The system's real job is to keep late ideas cheap.

Choosing is authorship

Thirteen prompts were a pick between options the machine had laid out. Which bed. Which singer: three voices became a wavetable singer because that one won the audition. Which voices for the conversation. Which master, A or B. Which face. Asking for a few quick versions and choosing between them is faster than describing what you want, and what you choose is what you mean.

Two sessions, one conductor

The two sessions sent each other five messages. The first final mix came back with the Italian buried by almost thirteen decibels. Nobody heard it first. It was measured, by the session that did not make the mix. The cause was a units mismatch between the two repos. Verification across a boundary catches what the maker cannot.

It worked the other way too. On one pass I could not hear her. A measurement said she sat two and a half decibels over the music. A minute later my ears agreed.

The second order

Some of this was recorded and never looked at until now.

The leverage is not in the typing. About seven words came back for every word I wrote, and two and a half lines of code. But the scarce thing was the four hours between turns. More than half the day was listening, watching and judging. The prompts are the small part.

Direction is layering, not a brief. 59 of 93 prompts landed while the machine was in the middle of something. Not a brief and then a wait. A note, then another note on top, while it plays. The same way the music was made.

The uncertainty became the structure. The music started with no genre and no clear goal, and that was said out loud. So it was built in layers from the ground up: sound, then music, then feeling. Those are the actual names of the ChipForge files. Not knowing turned into the method.

The standing rules did quiet work. Every voice went through the same tamer chain. Every measurement used the same ruler in both repos. Nothing gets published without my go. Those rules came from earlier films and were applied without being asked for. That is the part of the system that learns.

One gate is still open. Every number in this post can be checked by a machine except one: whether the song moves anyone. That gate is yours.

The film in numbers

Runtime	3:28 with bookends
Key	D major, a Lydian colour on G
Tempo	about 122 beats a minute, every bar its own tempo
The aria	three voices fused into one wavetable singer, five Latin words
The conversation	fourteen Neapolitan lines, OG Bobby and Roberto
Sync	every measured word and line within half a frame
The house	live numbers in perspective, one unison at 739.99 Hz
Master	-16 LUFS
Picture	1280x720 at 24 frames a second

Made on a laptop

Python and PIL for the picture. An original ChipForge score generated from numpy. No GPU, no subscriptions, no stock footage.

More from Napkin Films

• GLOW JACKAL. Bobby becomes Anubis on a wall of bass.
• SAUL, SAUL. A cathedral, and a voice from the light.
• The Language of Sound, Season Three. The science under the music.

License

This film is licensed CC BY 4.0. Remix it, repost it, drop it into your own thing. Credit "Napkin Films / Organic Arts LLC" and link CC BY 4.0.

Engine code (Napkin Films, ChipForge) is GPL-3.0-or-later. ElevenLabs voice audio is licensed content and is not redistributed. With words of William Blake, Blaise Pascal and Isadora Duncan; after Gregory Bateson.