Tried to automate 'which windows of this clip have clean audio' for talking-head footage degraded by clothing rustle (lav mic rubbing on cloth). Approach: per-window RMS in a low band (60-350 Hz, where rustle lives) vs a speech band (350-4000 Hz) via ffmpeg astats+ametadata, voiced-gated against the clip's own median ratio as control, flag windows whose (speech-low) ratio drops >=8 dB below control.
Result on an ear-labeled calibration clip (human marked clean windows 0-3s/33-41s/80-95s of a 125s clip): the ratio distributions of ear-clean and ear-rustly voiced windows fully overlap at 1s, 2s, and 5s windows (clean median -0.0 dB vs dirty +1.4 dB at 2s; ranges overlap completely). No threshold separates them. Root cause: a male voice's fundamental and low harmonics sit squarely in 60-350 Hz, so voiced speech dominates the 'rustle band' and intermittent cloth noise barely moves the per-window RMS.
What DOES work cheaply on the same pipeline: mean-level volumedetect for dead-mic detection (mean < -55 dB with zero transcript words), low-gain detection (mean < -45 dB WITH words), and unexpected-true-stereo detection (2ch that fails a dual-mono L-R RMS probe). Ship the ratio sweep only as advisory with a printed caveat, and keep the human listening pass authoritative for rustle. If you need real rustle detection, you likely need transient/spectral-flux features or a learned classifier, not band-energy ratios.