Play the same room twice and the music isn't quite the same. That's not one long recording on shuffle — it's a system. A composer writes the score in separable parts, and the game decides which of them you hear, and when.
A game score is a decision tree that only becomes a piece while you play. The notes are written in advance; the arrangement is assembled live.
Game state — enemies near, health low, a mission turning — not the buttons you press.
Stems: separable layers, mixed and re-sequenced in real time.
The first trick is called vertical layering. The same eight bars keep playing; what changes is which stems are unmuted on top. Drive the three parameters below and listen — the drums kick in for combat and drop out when it’s over, exactly the way Hades does it.
Audio is off. Nothing plays until you press start — then drive the parameters and listen.
opens the pad and widens the chord
brings the drums in
adds the bright lead
Active state: Explore — decided by the drivers, the way an engine reads game variables.
Pad and bass settle into a steady, unhurried groove. You can hear: pad, bass and lead.
Every sound here is generated in your browser from oscillators and noise — there are no audio files on this page, and nothing from any game.
Layering changes how thick the music is. The other technique changes which music plays at all. It cuts a piece into segments and jumps between them — but never mid-phrase.
Each segment carries an entry cue and an exit cue. When the game asks for combat halfway through a bar, the change doesn’t fire immediately. By default it waits for the current segment’s exit cue, so the switch lands on a musical boundary instead of cutting a note in half.
In a tool like Wwise that waiting is spelled out in a transition matrix — a rule for every move from one state to another. A death might be set to interrupt at once; a victory might be told to let the current phrase finish first. Either way the change is quantised to the beat or the bar, which is why runtime music sounds performed rather than spliced.
Stingers are the punctuation on top: a cymbal swell or a horn hit, held back a fraction so it lands on the next beat. Press the stinger in the demo above and you’ll hear the bed duck under it for a moment — which is the subject of the next section.
A score that’s always there has a problem: it can drown out the line of dialogue you actually need to hear. Two techniques keep the mix readable.
The voice bus reaches over and pulls the music bus down — the way a DJ’s talk-over dips the record. The trick is in the timing: typically fast to duck (roughly 50–200 ms) and slow to release (closer to 500–1500 ms), so the music sinks out of the way and then eases back without pumping.
High dynamic range mixing takes the opposite approach: author sounds across a huge range of loudness, then slide a window over it. When a gunshot fires it sits at the top of the window and the rustle of leaves below simply drops out — and returns the instant the shot stops. No one decided the leaves mattered less; the system worked it out from loudness.
None of this is theory. Here are three soundtracks you may already know, and the technique each one leans on hardest.
Vertical layering: the drum stem kicks in during combat and turns off when the room is cleared, while a semi-random pick of guitar and bass stems keeps each chamber fresh.
Four pieces for Tartarus, three each for Asphodel and Elysium; bosses advance to a harder section.
An open-world score built from eleven stems in four-to-five-minute loops, no longer locked to one key or tempo — steered in real time by an internal AI Rockstar calls the Gunfight Conductor.
The first Red Dead managed five stems, all sharing key and tempo.
Compose a piece that stands on its own, then take it apart into intro, loop, and outro chunks the game recombines — adaptive scoring built before off-the-shelf middleware existed.
Bungie wrote their own tools; no system in 1996 could do what they wanted.
Studios rarely build the music engine themselves. Two packages do the job for almost everyone, and they split along a familiar line — power versus ease.
The most elaborate music system in the field.
A DAW-like timeline that composers take to quickly.
The demo in § I is the same idea, stripped to its bones. Each stem is a single gain node; a small scheduler quantises new layers to the beat; and nothing makes a sound until you ask it to, because the browser starts every audio context suspended. It’s a hand-built miniature of what Wwise and FMOD do at scale.