hyperframes-audio

HyperFrames Audio

A mix is a set of relationships, not a stack of processors. Two tracks that each sound right alone can be unlistenable together, and the fix is almost never "turn one down" — it is finding what they are fighting over and giving it to whichever one needs it. Every tool here exists to express one of those relationships.

Effects live on the element as data-fx-chain, and preview and render run the same Web Audio graph — the studio in a live context, the engine in an offline one inside the browser it already drives. There is one implementation of each effect, so what you hear while scrubbing is what gets written. You never tune twice.

Three attributes carry everything, all on the audio/video element itself:

Exact JSON for each, and the rules a lane must satisfy: references/attributes.md. Every effect with its parameters, ranges and units: references/fx-registry.md. How to work out what is wrong with a file you cannot hear: references/diagnosis.md. Presets, named jobs and one-knob profiles, plus a symptom-to-fix table: references/presets.md — read that before hand-building a chain, because one of the presets or named jobs usually already names the problem.

How it fits together

Two authoring surfaces write those attributes; two runtimes read them through the same builders. That shared middle is why preview predicts the render.

flowchart TB
  voice["voice track<br/>media file"]
  bed["music bed<br/>media file"]

  subgraph AUTHOR["Authoring — the only things that write attributes"]
    panel["Studio<br/>Voiceover carve control"]
    script["scripts/carve.mjs<br/>detects the pair, dynamic by default"]
    analysis["core/audioCarve.ts<br/>carveProfile · analyseCarveBands<br/>analyseCarveDuck · analyseCarveDynamics"]
    panel --> analysis
    script --> analysis
  end

  voice --> analysis
  bed --> analysis

  subgraph ATTRS["Written onto the bed element"]
    carveAttr["data-fx-carve<br/>source · strength · dynamic"]
    chainAttr["data-fx-chain<br/>peaking xN + gain, tagged fromCarve"]
    autoAttr["data-automation<br/>a lane per carved parameter"]
  end

  analysis --> carveAttr
  analysis --> chainAttr
  analysis --> autoAttr

  subgraph SHARED["One implementation, read by both"]
    build["audioFxGraph.ts · buildFxChain"]
    sched["audioFxAutomation.ts · scheduleChainAutomation"]
  end

  chainAttr --> build
  autoAttr --> sched

  build --> preview["Preview<br/>live AudioContext<br/>attachElementFxChain"]
  sched --> preview
  build --> render["Render<br/>OfflineAudioContext in the headless browser<br/>applyAudioFxChain"]
  sched --> render

  preview --> heard["what you hear while scrubbing"]
  render --> wav["processed WAV<br/>+ chainTailSeconds so the mix lets the tail through"]
  wav --> mix["engine · audioMixer<br/>volume lane baked into the PCM here, not in the graph"]
  mix --> out["the rendered mix"]

  edit["editing the attribute mid-playback"] -.->|MutationObserver| preview

The carve's own settings are never read at playback — the chain and lanes it produced are what play. data-fx-carve exists so strength can be changed on an existing carve instead of guessed back out of the filters.

Inside a carved bed the signal runs through the dips first, then the level match, then anything you built yourself — which is why a limiter you add still acts as the last ceiling:

flowchart LR
  src["decoded bed"] --> p1["peaking<br/>400 Hz"]
  p1 --> p2["peaking<br/>1 kHz"]
  p2 --> p3["peaking<br/>1.6 kHz"]
  p3 --> g["gain<br/>level match"]
  g --> hand["your own effects<br/>e.g. limiter"]
  hand --> dest["track gain, then out"]

  l1["lane fx.n1.gain"] -.->|"envelope of the voice's<br/>level in that band"| p1
  l4["lane fx.n4.gain"] -.->|"how far the bed<br/>ducks overall"| g

A static carve is the same graph with fixed values and no lanes at all.

First, work out what is wrong

The table below starts from "it sounds boomy" — which presumes somebody already listened and said so. Handed a file and "fix this", you have no such sentence and you cannot listen, so you have to measure. One rule governs all of it:

The absolute spectrum of a single unknown voice cannot be diagnosed. Formants are ±10 dB, fundamentals run 85–255 Hz, and sentences decline 5–6 dB as they end. Every one of those reads as a defect on its own, and every one of them is the speaker.

So compare, and compare against something inside the same file: the clean original if it exists, otherwise the pauses — whatever is audible in a gap is additive, and the gap's spectrum is the channel rather than the voice. Comparing against a published average spectrum or a synthesised control voice does not work: two speakers differ by more than most defects, and both wrong answers in the evaluation behind this guidance came from exactly that.

When there is no original and no usable silence, a static tonal defect is genuinely under-determined. Say so and offer the readings that fit, rather than picking one and building a chain on it.

Commands, traps and worked recipes: references/diagnosis.md. Read it before diagnosing a file nobody has described.

Start from the symptom

Once you know the band and the kind, name what is wrong with the audio. Most bad audio is one or two of these, and each has a shipped answer:

Full catalogue, what each preset contains, the band vocabulary, and what is deliberately NOT covered (de-essing, noise removal, tone match): references/presets.md.

Subtract before you add, level after you filter, relationships after level, character and ceiling last. Each step changes what the next one hears — a compressor set before a high-pass spends its time chasing rumble.

Reach for a family by the problem, not the name

Filters (highpass, lowpass, peaking, lowshelf, highshelf) decide which frequencies a track is allowed to occupy. This is the first tool for two sources colliding, because collisions happen in bands: a bed and a voice both want 1–3 kHz, and taking that from the bed costs the bed far less than turning the whole thing down costs the mix. A high-pass on a voice is the standard fix for rumble; a low-pass darkens or muffles deliberately.

Dynamics (gain, compressor, limiter, gate) decide how a track's level behaves over time. Compression narrows the distance between loud and quiet so the quiet parts can come up. A limiter is a ceiling — it does not shape anything, it guarantees nothing gets past. A gate removes what is below a threshold, which is how you silence room tone between phrases. gain is a plain level stage, and it is what an automation lane rides when a track has to move out of the way.

Nonlinear (saturate, bitcrush) changes the waveform's shape, which adds harmonics that were not there. Reach for it when a track needs character or grit rather than correction — and remember it is generative: it makes a thin source denser, not cleaner.

Time (delay, reverb, chorus, phaser) puts a track in a space or gives it width. These are the ones that most easily wreck a mix, because a tail or a detuned copy occupies the same room a voice needs. Use them on the thing that should sit behind something else, and keep the wet amount lower than sounds right in isolation.

The chain is serial: each effect processes what the one before it produced. So corrective filtering goes early, character in the middle, and a limiter last where it can actually act as a ceiling.

Voiceover carve

The problem it solves. A music bed under a voice makes the voice hard to follow. The reflex is to duck the whole bed, which works and costs the bed all of its presence — the music goes limp for the entire voiceover. But the voice does not need the whole spectrum. It needs the few bands it actually occupies. Carve takes only those, and the bed keeps its low end and its top, so it is still music while the voice is still intelligible.

It is a relationship, not an effect. The settings live on the bed — the track that gets processed — and they name the voices to listen to, exactly as a sidechain compressor does: you select the track that gets quieter and pick what makes it quieter. Never put a carve on a voice track. A voice carved against itself is a bug, not a subtle mix choice.

Every voice, not one of them. sources is a list, because a bed usually runs under a whole sequence — a narrator, an interview answer, a second presenter. They are summed onto the bed's own clock before anything is measured (mixCarveSources), so one analysis covers all of them: the bands come from all the speech there is, and the envelopes rise wherever any of it is happening. Voices that never play while the bed does are left out; they cannot mask it.

One knob. strength is 0..1 and derives everything: how deep to cut, how many bands, how wide, how far to favour intelligibility over raw voice energy, how far the level may drop, how far under the voice to aim. Those six move together in any real mix — a gentle carve is a shallow cut in few bands with little ducking, a hard one is deeper in more bands with more — so they are one relationship written once, in carveProfile. Default is 0.25 — a 6 dB dip in three bands with 6 dB of level room, audible without sounding like a hole. At 0.5 the dip reaches 10 dB, which is where a carve starts being heard as an effect rather than as room for the voice; above that is deliberate territory for a loud bed under a quiet voice. 0 is spectral only — one band, no level match at all.

Carve by default. A bed playing under narration wants a carve; it is not a polish step to get to if there is time. Place both tracks, run the command below, listen. Skip it only when there is no narration for the music to sit under — a music video, a title card, a montage cut to the track.

It always follows the voice. There is no static mode: a fixed depth thins the bed through every pause, and once you have heard both there is no reason to want it. Every value becomes an envelope of the speech's own level — silence leaves the bed alone, a loud passage pushes the carve to full depth — written as ordinary automation, which is why the lanes show up in the timeline and can be edited afterwards.

Level matching is part of it. Spectral carving cannot fix a bed that is simply louder than the voice. So the carve also measures how far over the voice the bed sits and writes a gain stage: held at one value for a static carve, driven by an envelope for a dynamic one. That envelope releases slowly on purpose — music that snaps back to full the instant a word ends sounds like a machine doing it.

Running it. In Studio the carve is one module at the top of a track's effect rack — voice, strength, dynamic, and the analysis it produced, in one card. It is there whenever another track could be the voice, and a bed with exactly one candidate above it is carved by default, dynamically, at the default strength: that is what a bed under narration wants, and the module is where you change or switch it off. Several candidates leaves the picker waiting rather than guessing. Headless — which is the path when you are authoring a composition rather than editing one:

node <SKILL_DIR>/scripts/carve.mjs --comp index.html

That is the whole command. It finds the voice and the bed itself, carves dynamically at the default strength, and prints what it decided:

bed    music-bed (name looks like music)
voice  narration (only track left)
carve  strength 0.25 dynamic
bands  400Hz -6dB q1.4, 1000Hz -3dB q1.4, 1600Hz -3.17dB q1.4
level  216-point envelope, floor -6 dB

Name the tracks with --bed / --voice (repeatable) when the automatic choice is wrong, --strength to push it, --dry-run to see that report and write nothing.

How it picks the tracks. Names first, because that is what you already told it and the answer is explainable — classifyAudioName in core, the same classifier Studio's own picker uses, so the two cannot disagree. A track whose id or filename looks like music (music, bgm, bed, score…) is the bed; everything else that plays over it and is not SFX-shaped is a voice. Audio elements are preferred: video counts only when no audio track is left to be the voice, or every B-roll clip in the composition would read as somebody talking. It refuses when it cannot tell which track is the bed rather than carving the wrong one — typing one id is cheap.

Same analysis functions as the panel, so the result is identical. Needs ffmpeg on PATH and @hyperframes/core installed in the project (npm i -D @hyperframes/core) — the CLI inlines core rather than shipping it, so it cannot be borrowed from there.

What it writes is an ordinary chain of peaking filters plus a gain stage, tagged fromCarve. That tagging is the whole trick: a re-run replaces the previous carve and leaves every effect you built by hand — and every lane you drew by hand — exactly where it was. So re-carving at a new strength is safe and repeatable, and data-fx-carve exists so the settings can be read back rather than guessed from the filters.

Automation

A lane is a set of breakpoints on one parameter: {t, v} in clip-local seconds and the parameter's own units. Targets are volume for the track's level, or fx.<nodeId>.<param> for an effect's knob.

Only some parameters can be automated, and a lane on the others is silently inert. A knob is automatable when a Web Audio AudioParam backs it. The four worklet-based effects — compressor, limiter, gate, bitcrush — expose none at all, so no lane on any of their parameters will ever move: to make a compressor's behaviour change over time, automate a gain stage before it instead. references/fx-registry.md marks every parameter.

Verify

Almost no static gate covers the mix. The linter reads data-automation for exactly one conflict — audio_volume_double_automation, a volume lane on a track that also has a GSAP tween on volume, where the lane wins and the tween is ignored — and nothing validates the chain or the effect lanes at all. What enforces those is the render: a chain it cannot parse fails the whole mix rather than quietly writing the dry signal, because a mix that sounds plausible and is wrong is worse than a refusal. Preview is the opposite by design: an unreadable chain plays dry so the composition stays workable.

A lane pointing at a node the chain does not have is pruned on read, not an error — so a typo'd nodeId costs you the envelope silently. Read the ids back out of the chain rather than assuming what was minted.

Effects with a tail (reverb, delay) make the rendered track longer than its source, and the mix is told how much by the chain. So a bed with reverb no longer ends exactly at its data-duration; that is expected, not a bug.

Beyond that, a mix is verified by rendering and listening. For a carve: the voice should be legible without the bed sounding hollowed, and with dynamic the bed should come back up between phrases rather than staying flat. If the bed sounds notched rather than simply quieter under the voice, the strength is too high — that is the one failure mode with an obvious sound.