→ all versions of this pipeline
You marked three. They turned out to be three to four and a half seconds long — not the fractions of a second I had been treating — which is why nothing I did before touched them.
What your marks changed. Four of the nine windows had no noise in them at all, so two thirds of what I was processing was never the problem. The three real ones are ten times longer than the spans I had derived. Every previous attempt was carefully treating the wrong half-second.
It also corrects something I published: the profile page says these events last 96 to 832 ms. That was my localisation showing through, not the sound. They last 3 to 4.6 seconds.
What was done, and why it cannot whistle. One gain across the whole spectrum, lowered smoothly wherever there is no voice and returned to normal wherever there is. The metallic ringing came from cutting narrow frequency bands harder than their neighbours; with a single broadband gain there are no neighbours to ring, so that artefact is impossible by construction rather than by tuning.
Measured: the noise falls 7 to 13 dB, the speech inside the same spans moves 0.6 to 0.9 dB — under what anyone hears as a level change — the gain never rises, and everything outside the three spans is bit-for-bit identical.
Each starts 1.5 s early. Is the noise gone, and is the speech untouched? There is an option for metallic artefacts too — if that is still there I want to know rather than guess.