Cleanup tools change more than the noise. This check lines your original and the cleaned version up, sample by sample where it can, then reads the same moments in both, so what it reports is what changed in your take.
Noise between words is read in the pauses and measured against the voice, so a louder export alone doesn’t count as less noise. High frequencies are read only where the voice carries them, so hiss that was removed can’t pass for a duller voice.
Musical noise is the short, warbling tones heavy noise removal can leave in quiet parts. Gating is the background switching off between words while it stays under the voice. Smeared s sounds spread wider in time, or come out softer than the rest of the voice.
We tested each measurement on real studio speech run through cleanups we built and through a real AI denoiser, and the check shows only the ones that raised no false alarms in those tests. Two we tried, cut word endings and a level evened out over time, weren’t reliable enough, so the check doesn’t report them.