Audio design

1,587 words

Music, effects, and the problem of silence — which in a game set mostly in vacuum is the central design question rather than an aesthetic one.


The silence problem

Space is silent. A game that honors that is authentic and, handled badly, feels broken; a game that ignores it sounds like every other space game.

The resolution: everything you hear is something you could plausibly hear. Every sound in the game is one of four kinds, and the player is never asked to accept a sound with no source.

Kind Heard through Examples
Structural The hull, by conduction Thrusters, the reactor, the drive spinning up, an impact
Atmospheric Air, when there is any Wind on a surface, the airlock cycling, atmospheric entry
Interior Your own suit or the cockpit Breathing, gauges, switches, fabric, footfalls
Interface The ship's own annunciators Warnings, scan completion, lock tones

Vacuum is genuinely silent, and the transitions across that boundary are the most important audio moments in the game: the airlock cycling down, the wind dying, and being left with your own breathing.

Structural conduction is the trick that makes it work. You hear your own drive because it is bolted to the hull you are sitting in. You do not hear the other ship's drive at all — you hear it only through your instruments, which is why the contacts panel has a sound and the ship does not.


Music

Two rooms, two answers. Flight is where the drive through the hull is the score, and music there is sparse and mostly absent. The Planetarium is a place to sit under the sky for an hour with nothing in it that makes a sound, and an hour of nothing reads as broken — so it plays a soundtrack (ADR-0059). The cinema plays its own scenes' tracks and nothing else.

In flight: sparse, ambient, and mostly absent

Long stretches with none, which is what makes the moments with it land.

Game state Music
Under burn The drive through the hull — low, industrial, felt more than heard. Under it, sparse tonal pads. The dominant state.
The flip Everything stops. Four seconds of hull noise and breathing, then the drive relights. The score does not fill it.
Arrival in a new system A single motif on the discovery scan resolve. Eight seconds. Then silence.
Surface, first egress on a new body A low sustained drone; texture, not melody
First discovery confirmed One chord. Not a fanfare.
Combat Percussive, low, and it fades out as soon as the threat resolves — including when the player runs
Anything else Nothing

Adaptive triggers are on state, not on timers: entering an atmosphere, falling below a heat threshold, acquiring a lock, completing a scan. In flight there is no looping combat track and no background music bed.

Reference: the ambient work in Elite Dangerous (Erasmus Talbot) for the in-space register, and Outer Wilds for the discipline of using very little. Not No Man's Sky — 65daysofstatic's score is excellent and it is continuously present, which is the opposite decision.

In the Planetarium: a soundtrack

Fifty-two full tracks, a little over four hours, in three categories: Ambient (Stellardrone, Chris Zabriskie, Lee Rosevere and others), Lofi (HoliznaCC0, Purrple Cat), and Signals — radio from spacecraft, made audible, which is not music and starts switched off so that somebody chooses to hear it between the tracks rather than meets it there.

Behavior
Shuffle On by default, and a bag rather than a die: every enabled track plays once before any plays twice, and a new cycle never opens on the last track.
Transitions An equal-power crossfade, six seconds by default and up to twelve, timed from each track's measured silence rather than its file length.
The credit A card for six seconds as each track starts — title, artist, release, license — announced politely; the Soundtrack panel brings it back.
Under a scene Faded out over a second and paused in place: a cutscene with a track of its own owns the room, and the music comes back where it was when the scene ends.
Under the guide Held 12 dB down while a guide session is live — a voice over music held well down, rather than music stopping for every exchange — and back up slowly when it ends.
Hidden tab Faded out, unless the listener asked for background play.
Levels Everything, then music, on a squared curve so halfway sounds like halfway; mono for one earbud.
Offline A track streams by range; a kept track plays without a connection, and keeping one is the listener's choice, with its size beside the switch.

It does not react to navigation yet. The player exposes cue() — a category or a track, at a point, over a fade — for an arrival to call when arrivals have a sound.

Only what anyone may share. Every track is CC0, public domain, or CC BY 3.0 or 4.0, because the repository is public and the site is static, and either one hands the files to anybody who asks. Every track is held to −18 LUFS integrated under a −1 dBTP ceiling by gain alone, so a shuffle across categories never jumps in level and nothing is limited to reach a number. The credits are at /credits, in each artist's own words.


Effects

Principle
Every mechanic has a sound before it has a UI element A player should know the heat is climbing before they look at the gauge
Warnings are pitched, not loud Three annunciator tones — advisory, caution, critical — distinguished by pitch and repetition rate, never by volume
Physical objects have mass in the mix A sample core set down sounds heavy; in low gravity it sounds heavy and slow
The suit is always audible Breathing, at a rate that tracks exertion and oxygen. It is the on-foot layer's heartbeat and it must never loop audibly.
Nothing is stereo-panned that should not be Sound conducted through the hull comes from everywhere, because it does

Voice

Flight and Survey correspondence have no performed narrator. The Planetarium's optional Guide is an explicit exception.

This is a scope decision and a tonal one, and they agree. All correspondence is text. The Survey communicates by message, not by radio, which is both cheaper and more consistent with a setting where the nearest other person is usually several light-years away.

The Planetarium guide performs an explanation when the visitor asks for one or starts a tour. Its delivery can vary in rhythm and emphasis, and its silence gives the view time to register. Voice starts with a deliberate action and an AI voice disclosure. Muting the guide preserves readable explanations; muting the microphone and pausing the tour remain separate controls.

The guide speaks in one voice: GPT Live listens and talks, and delegates the thinking to a backend model, so a tour and a conversation are the same voice with no clip to hand off to (ADR-0042). Speaking interrupts the guide, and the microphone remains available. Because there is no separate speech source, the only honest signal that a stop has been heard is the voice the visitor is already hearing: the browser reads the level of the remote audio track and treats a stop as spoken once the words have played and gone quiet. Provider access, pronunciation, device behavior, and listening quality remain measured acceptance work, not properties established by a model name.

The ship's annunciator has twelve fixed synthesized strings.

The annunciator is an instrument and sounds like a machine. Its fixed delivery is independent of the Planetarium guide's performance. Real aircraft do this for the same reason it is wanted here: under load, when you cannot look at a gauge, a spoken word is the fastest channel there is.

The twelve: life support, hull breach, thermal critical, fuel reserve, drive offline, reactor overload, pull up, terrain, collision, lock warning, oxygen, pressure. Flat, unhurried, and identical every time.


Budget

The whole audio requirement for flight is roughly 120 effects, 8 ambient beds, 6 music cues, and 12 annunciator tones, and it is deliberately sized to be one person's work over a few weeks rather than a discipline. It is met by recordings under the soundtrack's four licenses plus targeted recording — not by a licensed library. Every library surveyed (Sonniss, Pixabay, ZapSplat, Mixkit, Uppbeat and the rest) permits use inside a production and forbids redistributing its files as files, which a public repository and a static site both do (spike 6).

Procedural audio is the leverage. Thruster and drive sound should be synthesized from actual thrust and RPM state rather than sample-triggered, which is both cheaper in assets and dramatically better in a game where thrust is continuously variable. The Web Audio API is well-suited to it and this is one place the browser is an advantage rather than a constraint.


  • onfoot — breathing, and the vacuum boundary
  • world — why there is no swelling score on arrival
  • ADR-0059 — the soundtrack, its licenses and its delivery
  • technical — Web Audio budget