Adaptive Music with AI

Layering and resequencing only work when tempo, key and bar lines are shared

Related

StemsTempo

The problem

Adaptive music has to answer two questions at once, which layers are heard now, and where the next section enters. Ask AI for four intensity versions of one piece and they usually disagree, two beats apart in tempo, a semitone apart in key, built on different chord changes, and the crossfade exposes all of it.

How it works

Vertical layering depends on a shared tempo, key and harmonic grid, with each layer changing only density and register. Horizontal resequencing depends on bar lines, so switches must land at the end of a phrase or on a strong beat, ideally prepared half a bar early. Generation models do not respect these constraints on their own; they follow their own harmonic logic, so the constraints go into every prompt or into MIDI edits afterwards. The more layers there are, the simpler the shared skeleton must be, or each addition has to be realigned.

Practice

Build a four to eight bar skeleton of harmony and rhythm first, fixing key, tempo and the chord per bar. Generate each layer against that skeleton, or simply reuse one MIDI part with different orchestration. Label BPM and start bar on every export. Quantize resequencing to the bar in the engine and keep fades under half a beat. Test by opening all layers together and then muting them one at a time, listening for phase or key clashes. Write transitions as their own two bars instead of expecting a fade to fix everything. Check the skeleton is unchanged before each pass.

What to listen for

In Bolero, how instruments enter one by one and thicken the texture without changing the harmony; in the Bach chorale, how four voices give clean switch points at cadences; in The Blue Danube, how the three-beat phrases close every four bars; in loop-house-in-a-forest, how a loop is built in layers. These are the structures an adaptive system has to reproduce.

Sources

• 自适应音乐的纵向分层与横向重排，依游戏音频通识核对，未抄原文
• Vertical layering and horizontal resequencing, checked against general game audio practice, not copied

Used for fact-checking only; all prose is original.