→ all versions of this pipeline
Five attempts to remove the handling noise by replacing it failed. The answer turns out to be not to remove it at all — but to lower it to the level of the room around it, leaving the original audio in place.
What changed. Every earlier attempt cut the noise out and pasted something else in. That needs the span to be exactly right, needs a donor matched to the room, and leaves two seams — and it went nought for five.
This instead caps each frequency band inside the span at the level that band has in a pause nearby, and keeps the original phase. The recording stays; only the excess comes down. Nothing is synthesised, nothing is pasted, and there is no seam to hide. It is the same operation a dialogue editor reaches for, and the professional tools call it attenuate rather than repair for exactly this reason.
Checked before it touched your audio. On a test signal built to your measured profile: a +22 dB event came down to +1.6 dB, speech 50 ms away changed by −380 dB — which is to say not at all — and with no spans marked the output was bit-identical to the input. The gain can never exceed 1.0 by construction, so the process can only ever turn things down.
On the real episode: 0.95% of the file touched, and everything outside those spans transparent to twenty-three decimal places.
Two things I had wrong, which the tests caught. I first wrote that an over-wide span would be harmless. It is not — speech caught inside one gets turned down with the noise, so the check that refuses any span containing a voice stays. And my first run used the episode's quietest moments as the reference, which pulled these spans 9 to 20 dB below their own surroundings — digging exactly the hole this method exists to avoid. The reference now comes from the pauses either side.
Each plays from two seconds before, so you hear the run-in either way. The two questions: is the noise gone, and is the speech untouched. A third worth listening for — does it sound unnaturally clean, as though a hole opened where the room should be.
Refused automatically: a voice sits inside or at the edge of these. Turning them down would turn the voice down too.