Reviewing interview recordings: verify Whisper pauses and speakers with audio math

When critiquing a recorded interview, run Whisper large-v3-turbo with --word_timestamps True, an initial prompt full of disfluencies ('Um, uh, like, so') and --carry_initial_prompt True; that keeps um/uh so filler rates are measurable. Whisper compresses long silences and drops lone 'um's, so confirm stalls with per-0.25s dBFS windows (numpy RMS on the 16 kHz WAV) and re-transcribe just that window: it exposed an 8.6 s stall that the main pass showed as a single 'Um'. Meeting-tool diarization mislabels short interjections; a crude per-word autocorrelation F0 separates two voices well enough to attribute 'Nice.' or 'Okay.'. If the speaker's project is public, compare what they said against its README and type definitions; that contrast was the most useful finding. Homebrew ffmpeg may lack drawtext, so put timestamps in frame filenames instead of overlays.