meeting-clip-anonymizer

Turn a recorded call into branded short clips while anonymizing the other participant (voice pitch-shifted, face hidden, PII bleeped). Use when repurposing Zoom/Meet/Fathom calls into YouTube/Shorts.

SKILL.md

GitHub

Meeting Clip Anonymizer

Repurpose a recorded call into polished clips without exposing the other person: their voice is pitch-shifted (anonymized), their face is replaced by a branded card whenever they're on screen, and names/figures are bleeped.

When to use

  • Repurposing coaching/sales/podcast calls into content where only YOU are the public face.
  • Any time you need a participant's words but must protect their identity.

The capabilities (in order of reusability)

  1. voice_anonymize.py (standalone, start here): deepen a speaker's voice on given intervals. Pure ffmpeg, duration-preserving, instant. python voice_anonymize.py in.mp4 out.mp4 --intervals 0:9,41:48 --pitch 0.82
  2. Full pipeline (pipeline/): silence-removal + face-hiding (branded bullet cards) + voice-anon + PII bleep + branded captions/circles/cards/music/zoom + cloud render.

Workflow (agent steps)

  1. Transcribe the recording to word-level timestamps ({words:[{word,start,end}]}) — e.g. OpenAI whisper-1. Save as transcript.json.
  2. Decide the clips: read the transcript, pick self-contained advice moments (start/end). Propose to the user for approval BEFORE rendering. Author each clip's editing to match the transcript — circle the KEY phrases as they're spoken, drop a card on each real takeaway. Do NOT sprinkle circles on a timer; that looks generic. ~1 meaningful animation per 5-7s.
  3. Find the other person's speech intervals (to anonymize): prefer speaker-view detection (the on-screen tile = who's talking; classify_speakers.py template-matches the name label). Diarization (pyannote) is a fallback but UNRELIABLE for two similar same-mic voices.
  4. List PII (names, locations, $ figures) and get the user's approval, then bleep (audio mute + caption mask).
  5. Run pipeline/render_clip.sh <clip> (cloud) or stop after the base cut and edit elsewhere.
  6. Deliver the clip + a set of YouTube title options.

Editing rules that matter (learned in production)

  • Captions (running subtitles) go ONLY over the other person's question / B-roll — never over your face. Your answer is told with circled phrases + cards.
  • Cards/circles only over YOUR face; never two text elements on screen at once; never an overlay card on top of the face-hiding branded card.
  • Voice: pitch < 1.0 (DEEPER, ~0.82), not higher. Shift on substantial spans where the other person is the active tile (>=0.8s); don't shift sub-second flicker tails (that's your voice).
  • Silence: cut hesitations/dead air; when merging adjacent segments, ONLY merge if truly adjacent — merging across a cut gap re-fills the silence (a real bug we hit).
  • Zoom: ONE gentle continuous drift per shot (e.g. 1.0->1.045), never pulsing in/out.
  • Run heavy assembly on a server, not the user's laptop (it pegs CPU).

Privacy

Always have consent to record and publish. Anonymization reduces but doesn't eliminate identifiability (story details, word choice). Redact specifics and get sign-off before publishing.