Music under a voice should get out of the way when the voice speaks and come back when it stops. That is ducking, and it is a measurement of the voice followed by a level curve on the music.
What it listens for
The voice file is read once in 10 ms steps. Every step louder than Voice above is speech. Runs of speech become stretches; gaps shorter than Hold are bridged, so a breath does not bring the music up.
The curve
For each stretch the music goes down over Attack, finishing as the voice starts, stays down by Duck by, and comes back over Release. When the next stretch starts before the release ends, the music stays down.
Levels
Voice and Music set each file’s level before the duck. The duck comes off the music’s level. The sum is not scaled back automatically; the peak is reported after the export, and the limiter is one handoff away.
Two files, one mix
The output is as long as the longer file and takes the higher sample rate and channel count of the two. A mono file is fed to both sides.