Two microphones in one room: the guest’s voice is in the host’s microphone, 12 to 30 dB down and a few milliseconds late, and it muddies the mix and doubles every effect. Given both recordings, this turns the other voice down in this one, wherever it is the only voice there.
How the bleed is learned
Both recordings are read once, one peak per 10 ms. Over the stretches where the other microphone speaks and this one does not, the difference between the two settles at one number, how far down the other voice sits here: that is learned, and shown after the export. It needs two seconds of the other voice alone.
Where it acts
A stretch is turned down when the other microphone is speaking, this microphone holds the other voice at the learned level, and this person is well under their own speech. Stretches closer than the hold join. Where both speak, this voice sits above the bleed and the stretch is left alone; so is a quiet aside under the other’s laugh.
Turn down by, attack, hold and release
How far the other voice drops, up to 40 dB; 18 dB takes it out of the way without a hole. Attack is how fast the drop comes in, hold how short a gap keeps it down, release how fast the level comes back.
Example
An interview with the host and guest on separate microphones. On twenty built pairs of readers at one level, with the other voice 18 to 25 dB down, 92 % of the other voice’s time was turned down, none of this voice’s, and under 2 % of the time both spoke; the learned level was within 0.6 dB of the built one.
Where it stops
With this voice only a few dB over the other’s bleed, the two cannot be told apart by level where they overlap, and the tool says so: match the two files’ levels first. Room reverb that outlasts the other voice is turned down with it only while the hold lasts.