Game audio: sources and how they map to Godot

The research behind the game-audio skill and the audio command, gathered on 2026-10-21. The sources are middleware manuals (FMOD, Wwise, Unreal), Godot 4.7.2's class reference and engine source, and the published loudness recommendations. None of the web pages states a licence, so they are summarised and linked here, not copied.

Sources

Source	What it gave
FMOD Studio manual 2.04, "Mixing"	Bus hierarchy (events into group buses into master), return buses fed by sends, VCAs as the settings-screen volume controls, snapshots, bus instance limits
FMOD Studio manual 2.04, "FMOD Studio Concepts" and the event macro, instrument and effect references	Parameters, logic markers and quantised transitions; multi and scatterer instruments; max instances, stealing (oldest, furthest, quietest, virtualise, none), cooldown and five priority levels; compressor and sidechain
FMOD Engine, "Managing resources in the Core API"	Virtual voices: a far or quiet voice keeps time silently and comes back as if never stopped; 32 to 128 real voices is typical
Learn FMOD	The free course list (Skill Tree, the Celeste project). The Skill Tree needs a sign-in and was not read
Wwise help: interactive music (music segments, switch and playlist containers, the transition matrix, stingers)	Exit at next bar, beat, grid or cue; sync to the entry cue or the same time; stinger cooldowns; pre-entry and post-exit for reverb tails
Wwise help: attenuation and mixing	Distance curves for volume and filters; obstruction (direct sound only) versus occlusion (reverb sends too); bus hierarchy, auto-ducking and HDR
Unreal Engine 5.8, "Audio Modulation Quick Start"	Control buses split into designer buses (ambience, foley, footsteps) and user buses (dialogue, music, SFX) driven by the settings menu; reductions add in dB
Godot 4.7 class reference and the 4.7.2 engine source	Everything in the Godot column below
Sony ASWG-R001 v1.10 (2013)	Loudness targets and how to measure them
IESD Mix Recommendations v03.02 (G.A.N.G., 2015)	The same console target from the industry body; a different mobile target
EBU R128 and ITU-R BS.1770	What integrated loudness, true peak and loudness range mean, and how they are gated
Joonas Turner's GDC talk on game feel with sound (gdc/game-feel-with-sound.md)	Action sounds: start on the hit, short first, one 5 to 20 ms impact layer, one recording and master chain per family, slight ducking under the player's actions; the source of `audio`'s `onsetMs` warning
Godot 4.7, "Sync the gameplay with audio and music"	The song clock: playback position plus time since the last mix minus output latency, discarding readings that go back; the system clock drifts over long songs
John Johanas, Hi-Fi Rush GDC talk (notes)	Landing each action's impact on the beat by changing its animation speed; rewarding on-beat play without punishing off-beat play
Guy Somberg, "Lessons Learned from a Decade of Audio Programming" (GDC, 2014 inferred; notes in gdc/decade-of-audio-programming.md)	Max within radius, a voice cap per spot that keeps every direction audible; integer voice handles; 6 dB as half or double
Dominic Vega, "Building a Mixing Sandbox for Just Cause 4" (GDC talk; notes)	Snapshots driven by game state and time instead of sidechains for priority (a sidechain on the lowest bus is nearly always on, so its sounds get mastered hot); a storm's mix changing over 15 and 25 s; ambience muted past five engaged enemies; the player's vehicle moved to a higher bus instead of boosted
bfxr2 (Stephen Lavelle, MIT; its synthesis after Thomas Vian's SfxrSynth, Apache-2.0; wavetables AKWF, CC0)	`sfx`: the Bfxr synth, its seven presets and its seeded random numbers, ported to C# (`ci/godot/tools/Sfxr.cs`)
Xbox Accessibility Guideline 105	Separate volume or mute for music, voice, effects, ambience and narration, and a mono option

Sound effects from a seed (`sfx`)

bfxr2's Bfxr synth is DrPetter's sfxr as extended by bfxr: an oscillator (square with a sweeping duty, saw, sine, triangle, noise, bit noise, tan, whistle, breaker and three AKWF wavetables), an attack, sustain-with-punch and decay envelope, pitch slides, vibrato, two pitch jumps, a repeat, low- and high-pass filters, a flanger, bit crush and a compressor, rendered 8 times per output sample at 44.1 kHz. Sfxr.cs translates Bfxr_DSP.js (engine 1.0.4), the presets in Bfxr.js (pickup_coin from its template, as bfxr's editor and library do) and the mulberry32 generator and FNV-1a seed that lib/bfxrlib.js installs as Math.random, so the kit and bfxr2 make the same sound from the same seed. riffwave.js (GPL-2) was not used: the WAV writer is the kit's own.

Checked in October 2026 with bfxr2's library under Node.js 22: for each of the seven presets at seeds 1, 12345 and 777, the seed, every preset parameter and every sample were identical (21 sounds, largest difference 0). The .bfxr files sfx writes, opened by the same library, rendered to within half a 16-bit step of the WAVs. Godot 4.7.2 imported them as 44.1 kHz mono AudioStreamWAV. tests/Test-Sfx.ps1 keeps the seed-1 length, RMS and three samples of each preset, so a change is caught without Node.js.

Untouched, bfxr's sounds peak between about -18 dBFS (a plain square) and 0 dBFS; sfx scales each to -PeakDb (-6). Measured with audio: the true peak of six of them was 1.2 to 2.3 dB above the sample peak (sharp edges overshoot between samples), so a peak of 0 dBFS would clip after conversion.
C# compiles in about 5 s the first time in a PowerShell session; a sound then takes about 15 ms.
-Vary is bfxr2 library's gentle mutation (each parameter, with even odds, moved by up to 5% of its range), not its editor's, which also re-rolls the pitch.

Middleware ideas in Godot

Middleware	Godot 4.7
Group bus	An `AudioServer` bus; every player's `Bus`
Return bus and send	A bus that other buses send to, holding a shared reverb; an `Area3D` with a reverb bus sends a room's sounds to it
VCA, Unreal user bus	The player-facing buses (Music, Voice, SFX, Ambience, UI) kept at 0 dB in the layout and set only by the settings menu
Snapshot	A set of bus volumes and enabled effects, faded with a `Tween`
Horizontal resequencing, transition matrix	`AudioStreamInteractive` and its transition table
Vertical remixing	`AudioStreamSynchronized`
Playlist container	`AudioStreamPlaylist`
Multi instrument, random container	`AudioStreamRandomizer`
Max instances and stealing	`MaxPolyphony` per player (oldest stolen); categories and priorities in your own code
Virtual voices	Not built in: do not start sounds out of range
Attenuation curve	`AttenuationModel`, `UnitSize`, `MaxDistance`, the attenuation filter
Obstruction and occlusion	Not built in: ray-cast and lower volume and filter cutoff
Sidechain ducking	`AudioEffectCompressor.Sidechain`
Master limiter	`AudioEffectHardLimiter`

Godot facts that are easy to get wrong

Checked against the 4.7.2 source or by running it.

A bus named in a player that does not exist falls back to Master at run time, with no error.
AudioStreamRandomizer.RandomPitch is a ratio: the pitch falls between 1/r and r, drawn log-uniformly. 1.05 is about ±5%. RandomVolumeOffsetDb of 2 means -2 to +2 dB. The default mode, RandomNoRepeats, never plays the same entry twice in a row.
MaxPolyphony defaults to 1. Past it, the oldest voice of that player is stopped with a 64-sample fade. There is no global voice limit and no priority system.
A position, volume or pitch set on a player just after play(), in the same frame, is heard from the first sample: 2D and 3D players start at the next physics tick. On the dummy driver (no sound device), a reused AudioStreamPlayer2D can play a short sound with its previous pan and volume; a new player or play(0.0001) avoids it. Measurements in gdc/decade-of-audio-programming.md.
AudioStreamInteractive quantises to NextBeat and NextBar only when the source clip has a BPM. Ogg Vorbis and MP3 streams carry Bpm, BeatCount and BarBeats; WAV has none. On a WAV clip the switch is immediate and the fade length is read as seconds instead of beats.
AudioStreamSynchronized changes a layer's volume immediately; tween it.
AudioStreamPlayer3D: inverse distance loses 6 dB per doubling from UnitSize, inverse square 12 dB. MaxDistance also scales the volume down linearly to silence. The attenuation filter is a high shelf at AttenuationFilterCutoffHz.
Web exports default to sample playback, in which bus effects, reverb buses and doppler do not run.
Movie Maker (--write-movie) records exactly what reaches the speakers, buses and effects included, at a fixed frame rate, so the result is the same on every run. It needs a renderer: with --headless it crashes. Under a virtual display with the OpenGL renderer it works. It writes .avi (MJPEG with PCM audio), a PNG sequence with a WAV beside it, or .ogv; there is no WAV-only writer.
A mono file played centred measures about 3 LU louder in the stereo mix than on its own. Measured on the kit's fixture, a -27.2 LUFS file gave -24.4 LUFS.
Godot 4.7.2 prints ERROR: 1 resources still in use at exit when a windowed run quits while a sound plays. The kit ignores that line.

Loudness

ASWG-R001 v1.10 recommends the following. Version 1.0 said -23 LUFS, which is why some articles quote that figure.

-24 LKFS ± 2 for titles on home platforms.
-18 LKFS ± 2 for titles on portable platforms.
True peak at most -1 dBTP.
The mix is measured "as a whole", not dialogue, effects or music separately.
Measurement covers at least 30 minutes of representative play.

IESD agrees for consoles (-24 ± 2 LU, -1 dBTP) and gives -16 ± 2 LU for mobile, with a loudness range of 10 to 15 LU. LUFS and LKFS are the same unit.

The audio command measures with ffmpeg's loudnorm filter, which reports input_i (integrated), input_tp (true peak) and input_lra (range):

Integrated loudness is gated: everything under -70 LUFS is ignored, so silence reads as no value.
A clip shorter than the 0.4 s measurement block has no integrated loudness at all.
For short sound effects, check the peak, not the integrated loudness.

Figures with no official source

The FMOD and Wwise manuals give none of these figures, so the skill offers them only as starting points to set by ear:

How far to duck music under dialogue. Blogs say 6 to 12 dB; this kit suggests starting at 3 to 6 dB.
Attack and release times for ducking. Blogs say 10 to 80 ms and 250 to 700 ms.
How many takes a footstep needs. Blogs say 3 to 8; the kit asks for at least five.
How much to vary pitch. Blogs say a few percent to 10%.
How late an action sound may start. Turner's talk gives 5 to 20 ms as the length of the impact at the start of a sound, not as a delay limit. audio warns when an sfx or ui file's onsetMs (the first 10 ms window within 10 dB of the file's loudest) is over 20 ms; the number is the kit's choice, and -MaxOnsetMs changes it.

The FMOD compressor's own defaults are 2.5:1, 20 ms attack and 200 ms release.

Mono and voice budgets (measured October 2026)

Mono. Godot has no mono switch, but AudioEffectStereoEnhance with pan_pullout 0 removes the difference between the channels. tests/mono-fixture plays a 440 Hz tone at 0.5 through Master with the effect off and on and reads the output back with an AudioEffectCapture (RMS of each channel, on 4.7.2 .NET and standard): a left-only tone goes from 0.353/0 to 0.177/0.177, a centred one stays 0.353/0.353, and the same tone in opposite phase goes from 0.353/0.353 to 0/0. So each side is (L+R)/2: a hard-panned sound loses 6 dB per side (3 dB of power), and stereo effects that rely on phase (wide reverbs, some synth pads) thin out or vanish; listen to the mix in mono before release. With pan_pullout 0.5 the left-only tone measured 0.263/0.088, a partial narrowing.
Voice budgets. VoiceBudget in Example.Core: per category a cap (a new voice steals that category's oldest) and a priority (at the total cap, a new voice steals the oldest voice of the lowest priority below its own; equal or higher priorities are never taken), and protected categories that are never stolen and are refused when full. A 5,000-request storm in its tests never exceeds any cap and never steals dialogue.