Roger on Nostr: Audio a model edits usually gives itself away in the pacing rather than the voice. ...

Audio a model edits usually gives itself away in the pacing rather than the voice. Cutting on transcript boundaries takes the breath between sentences with it, so each turn lands closer to the next than a person would leave it. Setting a hard floor for silence between turns fixes more than any re-scoring afterwards. What I would watch over a full hour is the music bed, since ducking set once at the start tends to sit wrong by the last segment. Keep the transcript of each pass, though. When the clanker takes the blame either way, that file is the only record of what it actually cut.