Presets

Reach for one when it names the problem you actually have. Build by hand when none of them does — a preset applied because it was nearby is worse than three deliberate nodes.

Diagnose first: what to listen for, and what fixes it

Work from the symptom, not from the effect list. Most bad audio is one or two of these, and the fix is usually a job rather than a whole preset.

The band vocabulary these map onto — the same names the rack shows:

Order of operations

Diagnose in this order, because each step changes what the next one hears:

1. Subtract before you add. Cut rumble and mud first. A voice that sounds dull often has too much low-mid, not too little top — lifting the top of a muddy voice makes it muddy and harsh.
2. Level after you filter. A compressor reacts to whatever is loudest, and a rumble it can no longer see is a rumble it stops chasing.
3. Relationships after level. Carve a bed against a voice once the voice itself is settled, or the analysis measures a problem you are about to fix.
4. Character, then ceiling. Saturation and space go late; a limiter goes last, where it can actually act as a ceiling. Anything after it is not bounded by it.

Presets

Four families, listed in full below. Apply one and it appends — stacking a character preset onto an already-cleaned voice is a real thing to want. Re-applying one that is already present replaces its own nodes in place, because position in the chain is signal order.

Voice — make a real voice sound like its better self

voice-clean is the default answer to “fix this voiceover”. The other two are the same idea pushed in one direction: broadcast is denser and more forward, warm has body added rather than cut.

Repair — one problem, one node

Character — deliberate, not corrective

telephone, radio-am, megaphone, lofi-tape, pa-system (Tannoy), intercom, doofus-worble.

These are costumes. Each is a band restriction plus a resonance plus its own kind of dirt, and they are tuned to be distinguishable from one another — measured on a log sweep, no two sit closer than the signal itself. Do not stack two.

Space — put it somewhere

room-tight (presence without wash), room-natural (recorded somewhere rather than nowhere), hall (far back and big), slap-echo (one quick repeat), dub-throw (repeats trailing well behind).

Use these on whatever should sit behind something else, and keep the wet amount lower than sounds right in isolation — a tail occupies the room a voice needs.

The whole preset as one control

A preset’s nodes are wrapped in a wet/dry blend, so presetAmount (0..1) fades the entire thing in or out, and fx.preset.<id> is an automation target that ramps it over time. This is the only way to automate a preset as a unit: its nodes share no common parameter, and worklet effects (compressor, limiter, gate, bitcrush) expose no automatable parameters at all.

Jobs — the range IS the module

Five named peaking filters with the frequency already chosen. Picking the job is picking the range, which is what makes a single “how much” knob honest.

Each is an ordinary peaking node underneath — the frequency is a starting point, not a cage. Prefer a job to a bare peaking when one matches: it arrives already aimed, and the rack names it for the work rather than the mechanism.

Writing one by hand, carry the name in label — {"type":"peaking","id":"n2", "label":"Reduce Mud","params":{"frequency":250,"gain":-3,"q":1.2}}. The parameters alone are not the job. A chain with three unlabelled peaking nodes shows the author three identical rows, which is the exact problem jobs exist to dissolve.

Every job also ships inside a preset, at identical settings — that is where the five came from. boom-tame is Tame Boominess; harsh-tame is Soften Harshness; voice-clean contains Reduce Mud and Add Clarity; voice-broadcast contains Reduce Boxiness. So check what a preset already contains before adding a job on top of it, or the cut lands twice — voice-clean plus a Reduce Mud job is −6 dB at 250 Hz where −3 was meant. The rack shows the contained nodes by name once the preset is expanded, which is the fastest way to see it.

One-knob profiles

Five effects have no single parameter that can honestly be their face — a compressor’s threshold means nothing without its ratio. They get a derived control instead, 0..1, which sets several parameters together.

Evenness, Warmth and Space are level-matched — the make-up gain, the output trim and the dry leg move with the drive, so turning the knob up does not also turn the track up or down. Those figures were solved by measurement, not chosen: the compressor originally left a track 2.5 dB quieter at full evenness, and saturation’s trim ran the wrong way entirely.

Tightness and Crush are not level-matched, because neither has a trim to move — a gate only removes, and Crush’s mix is the effect itself rather than a make-up.

The chain stores the mechanism values, not the knob position; the knob is read back by inverting the curve. So hand-editing a parameter under a profile is allowed and will simply move the knob.

Measuring scripts, not presets

Two things measure the audio before they act, so they cannot be a fixed chain:

• Voiceover carve — analyses the voice and cuts the bed in the bands the voice occupies. The answer to “the music is fighting the voice”. See the carve section in SKILL.md.
• Even Out Levels (levellingResult) — measures the track’s own speaking windows and writes a gain envelope. Its target is the 80th percentile of that track, not an absolute level, so an already-even track is left alone. Use it over a compressor when the problem is passages drifting over a whole take rather than word-to-word dynamics.

Not covered by anything shipped

Name the gap rather than reaching for the nearest preset and calling it the thing — but then ship the honest fallback anyway, with its cost stated. An author who asked for a fix and got only an explanation has been told something true and handed nothing. Say what it is, say what it costs, apply it.

• De-essing. harsh-tame is a broad always-on cut centred a band too low, not a de-esser. A real one needs a detector faster than the analysis hop available here. Fallback: a narrow peaking cut in the Edge band — sweep 5–9 kHz to find where this voice actually spits, Q 3–4, −3 to −5 dB. It is always on, so it costs a little air on every word; that trade is usually worth it and is the author’s to reject.
• Tone matching one track to another. Fallback: the Tone EQ by hand, which is predictable in a way a match curve derived from two takes would not be.
• Noise removal. room-gate closes the gaps; the noise under speech is untouched. There is no fallback for hiss beneath the words — a source with audible hiss needs a better source, and saying so is the whole answer.