game.sound — the music, the mix and the clip library

Everything SENGOKU TRANCE makes a noise with. Three psytrance tracks sequenced to the exact BPM each stage declares, the shared one-shot library every other slice reaches for, the crowd bed, the UI kit, and the mix that keeps them out of each other's way.

local ST = require("game.core")
local S = ST.need("game.sound") -- nil while this slice is unbuilt; handle it

Exports

S.play and S.clip take a library name or an alias, and an unknown name is a quiet no-op — a slice that guesses cannot take the game down.

The library

Aliases that resolve to one of the above: hit kiwami shin ka perfect great good kuzure fail click hover confirm select accept back cancel pause start tick metronome cheer crowd roar drum gong flute win lose begin combo danger sweep boom.

ui_tick is the metronome the calibration card should tap against — a 60 ms wooden click with a hard front edge.

What answers on the bus

Nobody has to call anything. These fire on their own:

The judgement topics are answered directly as well as through sfx, so a hit sounds whether or not the conductor also emits one; a repeat of the same name inside its gap is dropped, so the two paths cannot flam.

The mix

Nothing sits at the same level as anything else. Master 0.90; a spoken line 1.00, the track 0.68, the crowd 0.30, and the field kit quietest of all, at 0.22–0.30 of the sfx channel, so a menu never argues with the music. Under a story.line the track drops to 42% and walks back up over 1.6 s.

The tracks — the BPM is the contract

The three records are sequenced, not generated: a bank of one-shots stamped onto a grid derived from the bpm in ST.STAGES, section by section, following the psytrance shape that table declares. That is why a chart authored in beats and a track rendered from beats cannot disagree.

Measured by decoding each finished record and finding the loudest sample within ±55 ms of every beat of the grid. The constant few milliseconds is the kick's own attack shape — its peak sits about 4 ms behind its onset — and the two halves agreeing to a third of a millisecond is what says the tempo does not drift.

Each record carries a generated, wholly un-rhythmic atmosphere under it — rain in the gorge, gunsmoke, the temple burning — looped seamlessly, read at two different points for the two sides so the field is decorrelated, and pulled down to a tenth for the drop.

Rebuilding the bank

Everything under /source/audio is rendered by bake.luau. Run it from an execute, never from a component or the scene entrypoint: an asset.create made from either writes to the runtime store instead of /zero/source and the clip does not survive a reload.

execute { file = "/zero/source/game/sound.module/bake.luau",
          args = { what = "library",
                   taikoSrc = "…", fluteSrc = "…", roarSrc = "…", bedSrc = "…" } }

execute { file = "/zero/source/game/sound.module/bake.luau",
          args = { what = "track", stage = "okehazama",
                   taikoSrc = "…", fluteSrc = "…", atmosSrc = "…" } }

The *Src arguments point at generated .soundClip folders and are all optional — a synthesised voice stands in for each one, so a failed generation never leaves the game silent. The noise is a seeded LCG, so a re-bake is byte-identical.

Notes

• mount() builds one deck, one crowd bed and a ten-slot one-shot pool under ST.anchor("Audio"), and sets the four channel levels. Nothing is synthesised at mount — the bank is on disk.
• Every clip reference is a static literal: a computed asset.resolve on a gameplay path cannot be statically pinned and breaks when the world is imported elsewhere.
• The crowd bed is codec = "pcm", because an Opus clip loses its first few milliseconds on decode and a bed that loops would click at the seam. Loop points do nothing on this build, so the bed is authored seamless end to end and looped by the Audio component's own looping field.
• One-shots are Opus, which keeps the whole library under a megabyte.