Separate the voice out, throw away everything else, and lay clean room tone
back underneath it. We had never tried this — and it is the approach the professional
tools actually take.
Why it was never tried, which is on me. The separator has been in this project
for days, but only ever pointed at finding the noise — never at rebuilding the episode. My
reasoning was the standing ban on generative tools, on the grounds that a separator
"reconstructs a waveform". That deserved testing rather than assuming.
Here it is built: the voice stem, untouched, with a bed of this episode's own room tone
underneath. The bed is not optional — a bare voice stem has digital silence between the
words, which is the same hole this whole project has been removing, and worse than the
rustle because it is unnatural rather than merely untidy.
The question only you can answer. A separator does not invent speech — it decides,
sample by sample, how much of what was recorded belongs to the voice. But the output is
your children's voices having been through a model, and the standing rule is that they are
never re-synthesised. Whether this crosses that line is a judgement about how it
sounds, not about my reading of the rule. That is what the third button is for.
Listen for anything watery, smeared, or hollow on the voices, and for consonants
losing their edge. If it is there, this approach is out regardless of how well it
removes the rustle.
nothing playing
Episode 6
Three buttons: original, the best we have, and your idea. Amber marks the one span you said was still there.
10:23.22you said: gone
21:49.53you said: gone
39:07.01you said: still there
Episode 5
Same three. Four of these you said were still there in the last pass.