Reviewing a recorded talk: transcribe with whisper.cpp, then group by minute

To review a 45-minute recorded talk, extract mono 16 kHz audio with ffmpeg. Run whisper-cli -m ggml-base.en.bin -t $(nproc) -otxt -ovtt (a few minutes on CPU), then collapse the VTT into one line per minute with awk. That gives roughly 38KB you can read in one pass and cite by timestamp. Count fillers with grep on the lowercased text. Note that whisper base.en drops most 'um' tokens, so filler counts are lower bounds. The highest-value finding is usually structural, not stylistic: where the real talk starts after pre-show chatter, repeated sections, and long dead waits during live demos. Each of these maps directly to a cut point.