Hybrid Human and AI Performance

With a human performance plus generated parts, decide who owns the beat before discussing tone.

Related

RubatoTempoDynamics

The problem

When a human performance meets generated parts, the first thing to fall apart is rarely tone. It is time. Players push, drag and breathe at the end of phrases. If the generated part is locked to a fixed BPM, two timelines pull against each other. Listeners cannot name the fault; they just say it does not sit together.

How it works

Ensemble timing follows one of two models. Either the soloist sets the pace and everyone follows, with tempo flexing by phrase, or the rhythm section holds the beat and the soloist floats above it. A generated part must belong clearly to one model. It cannot track the beat and the expression at the same time; choose one and supply the other through a tempo map. A tempo map records where each beat actually falls, which is closer to human time than a single BPM value.

Practice

Record the human first without a click, then extract a tempo map from that take. Drive the generated part from the map, not from a fixed tempo. Keep the player's ten to thirty milliseconds of push and drag instead of quantising to the grid. Set the generated part three to six decibels lower, give it the same room sound, and place both in one reverb. Check that breaths line up, especially at the end of sections and at turns. Finally, have the player listen. If it feels like being dragged, start again rather than hiding it with compression or delay.

What to listen for

In Mozart's piano trio, listen to who yields between piano and strings. In the slow movement of Beethoven's quartet, listen to four players breathing together. In Grieg's piano concerto, listen to soloist and orchestra trading the lead. In the soul groove at 86, listen to where the laid-back feel actually lands. Hybrid performance succeeds or fails in these details.

Sources

• 合奏时间基准与速度图的做法，依演奏与制作实践核对，未抄原文
• Ensemble timing and tempo maps, checked against general performance practice, no text copied

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