Features

I. Instrument physics

Zaraforte's most important instruments aren't descriptions of a sound. They're simulations of the mechanism that makes it, run sample by sample, so the tone, the attack and the way the sound changes with effort come out of the physics rather than being painted on.

• Bowed strings (violin, viola, cello, double bass): a string that sticks to the bow and slips, over a resonant body. Bow speed and pressure follow the dynamics, and the player's hand adds a vibrato that is never quite the same twice.
• Brass (horn, trumpet, cornet, trombone, tuba): a pair of lips buzzing on a tube, with a mouthpiece cup, a bore that brightens as it's pushed harder, wall losses and a bell. The trombone's slide glides.
• Reeds and air: a clarinet's reed beating on a cylinder; the four saxophones, the same reed on a cone; flute and piccolo as a jet of air across a pipe, where the breath's own turbulence is what starts the note.
• The piano: a hammer whose felt stiffens the harder it strikes (part of why a loud note is a brighter note), strings coupled through a bridge to a soundboard, and the pedals, half-pedalling and the soft pedal included, with undamped strings ringing in sympathy.
• The triangle: a steel rod ringing in its natural modes, sized like a real triangle and rolled by a player whose hand slows down on a bigger one.
• The pipe organ: ranks of pipes, each with the chiff of a pipe starting to speak and the gentle wander of a wind supply that's never perfectly steady, in historical temperaments as well as equal.

Other voices, like the organ's stops, the timpani and the bagpipes, are recipes: a careful description of what the instrument radiates. The best of them are measurements; the bagpipes were read harmonic by harmonic off a recording of massed pipers. Zaraforte keeps the two kinds distinct and says which is which.

II. Human performance, and the flock

A machine playing exactly on the beat sounds like a machine. Human ensembles are loose in a particular way: each player keeps nudging their timing toward everyone else's, a little late here, a little early there, and the group holds together without ever being perfectly together.

Zaraforte's players do the same. Their timing follows the model researchers measured in real string quartets (Wing and colleagues, 2014), each player correcting toward the others in proportion to how far apart they are, which is a one-dimensional flock. Tuned against real recordings, it reproduces the spread human players actually show: a few tens of milliseconds.

Inside each note there's more of the player: breath and tonguing on the winds, the scoop of lips settling onto a pitch, bow changes, and a vibrato made by a hand, not a metronome: its speed and depth shift a little every cycle.

III. Chains of cause, from ensemble to fingertip

Design principle: model the whole chain of cause, not just its result.

A note in an orchestra is the end of a long chain. The ensemble's shared sense of time moves a player; the player moves an arm, a wrist, a hand; the hand moves a bow, a beater or a key; that sets a string, a rod or a column of air vibrating inside an instrument with its own body; and the instrument sings into a room, where a microphone hears it.

Zaraforte builds each link as its own part and lets them push on each other. The dynamics in a score reach the bow's pressure, not just a volume knob. A triangle roll is a hand working in a corner of a steel rod, so a bigger triangle is rolled more slowly and played wider. A pedal lifts the dampers off strings that then ring along with the notes being played.

IV. A notation language of its own

Zaraforte's scores are plain text, compact enough to type a riff and complete enough to hold a fugue. It reads MusicXML and MIDI, so existing scores and transcriptions come straight in, and then it goes where those formats stop.

• Dynamics as curves. A crescendo is a line drawn over time, like automation in a recording studio, so it can swell across a note that's already sounding. Hairpins can bow; a sudden change is simply a jump.
• Each player's timing. A lane of small offsets, in milliseconds, per player. This is how the flock gets into a score: written down where anyone can read it, and the same every time.
• How the instrument is played. Choose the mechanism (lips, reed or air jet), change an organ's stops from a given beat on, set a triangle's size and how fast it's rolled.
• Where everyone sits. Left or right, near or far, and the room they're in.
• Gracenotes, and cells. Ornaments that borrow their time from the note before, and named musical cells replayed by a form, the way a groove really repeats.

A taste, from a Bach fugue on trombones and a triangle roll:

voice tbn1 inst=trombone model=lip x=-0.5 far=0.5
voice triangle inst=triangle size=9.3

dyn tbn1 @0 mp < @17 mf   @38 mf < @46 f
time tbn1 @24 +11 @26 -3 @40 -12
register triangle @72 roll=7.3

• model=lip: this trombone is the lips-on-a-tube simulation; x and far place it on the stage.
• dyn: mezzo-piano growing to mezzo-forte by beat 17, then on up to forte.
• time: from beat 24 this player comes in 11 ms late, then 3 ms early, then 12 ms early.
• register: at beat 72 the triangle, a nine-inch one, starts a slow, grand roll.

V. Score analysis

Plenty of great music exists only as an engraved page. Zaraforte starts from the page: optical music recognition reads the engraved score into notes, dynamics and articulations, and then several correction passes go after what the reader got wrong, because on a complicated orchestral score it always gets something wrong.

The sharpest of those passes is mechanical. A harmony check scans the whole score for notes that don't fit the harmony sounding around them, and finds in minutes what listening finds one mistake at a time: a clef change that landed a bar late and put a bassoon a fifth out, or a horn part read in the wrong key. Aligning the result against a real recording catches the rest, like a repeat the reader skipped.

VI. Sonic analysis

Before Zaraforte builds an instrument or a performance, it measures the real thing. From a recording it reads tempo and how it breathes, tuning to the cent, key, the harmonic makeup of single notes across an instrument's range, how loudness and brightness change with effort, when each player in an ensemble arrived, and loudness over time.

Then it measures its own takes with the very same tools, side by side with the reference. That loop keeps it honest. It's how the piano was fitted note by note against recordings of a Steinway grand, and how our Also sprach Zarathustra was measured against a real orchestra's dynamic range, and learned to breathe wider. It once showed that a mysterious "note" in a recording was really the upper ring of a kettledrum.

VII. Room and microphone acoustics

Every player sits somewhere, and the sound is worked out from that seat: direction, the fall of level with distance, the balance between direct sound and the room's reverberation, the high frequencies that air absorbs over distance, and the moment the sound arrives. Orchestral seating is built in.

The rooms are synthesized too, from a stone church to a concert hall, a small club, a tiny office concert and a tiled subway station. Spot microphones can be placed close to a player, the way studios do, because at a few metres a microphone in a small room hears more room than instrument.

VIII. Mixing and mastering

After the room comes the desk. Each voice and the final mix can run through a signal chain of compression, equalisation, convolution and algorithmic reverb (plate included), drive, delay, chorus, a rotating Leslie speaker and more, down to deliberate noise and wear: tape hiss, wow and flutter, vinyl crackle, AM and FM radio.

The master is finished like a real one, with limiting, stereo width, dither and a loudness target, and loudness is measured to the broadcast standards (ITU-R BS.1770 and EBU R128), moment to moment as well as overall.

IX. Deterministic performances

Design principle: the same starting conditions produce exactly the same performance, down to the exact waveform.

Zaraforte's performances are full of variation, and none of it is left to chance. Every wobble of vibrato and every millisecond a player is late comes from the score or from a seeded sequence, so a take can sound human and still be reproduced byte for byte from its score, its seed and the version of the machine that played it.

That's more than tidiness. It means any change to the machine can be tested by rendering before and after and comparing the files exactly, and when two takes are compared by ear, the only difference is the one we meant to make. The principle is checked, not assumed: automatic checks replay pieces and compare them, and when a piece has failed to hold still, the cause was found and a test added so it can't slip back.

X. A human ear has the last word

Measurements say whether something changed. They don't say whether it sounds like music. Every instrument and every performance goes through rounds of careful listening, and the listener's notes (the instruments swell, it sounds a little rushed, that note is mistuned) are written down word for word and turned into the next measurement. The ear finds what's wrong; the numbers find where, and whether it's fixed.

XI. Every take comes with a picture of its sound

Each performance is rendered at 48 kHz and delivered with a picture of its own sound: the whole performance's spectrogram as one still image, a playhead that travels along it as the music plays, and lossless audio underneath. Every take is kept and stamped with the score, the seed and the machine version that made it, so any take can be made again exactly.