voice-audio

Files (1)

• SKILL.md

SKILL.md (first lines)

# Voice & audio

You cannot listen. Cast on evidence, not hope.

## Casting protocol
1. Enumerate the server's audio tools and every voice/style/quality
   parameter they expose.
2. Shortlist 2–3 voices whose descriptions fit the brand: natural,
   confident, a film narrator — not a screen reader. Generate the SAME
   one-line test with each.
3. ffprobe each test: duration sane for the line (~2 words/second),
   healthy loudness, no clipping. Commit to ONE voice.
4. Write for the ear: <=30 words across 15 seconds, short sentences,
   punctuation shapes delivery. Never read a spec sheet.
5. Walking away is a valid cast: if no voice meets the bar, go music-led —
   a strong track beats a mediocre read every time.

## The mix
- Voiceover clearly on top: duck music 6-9 dB under it
  (sidechain or keyframed volume).
- Normalize the final mix: `-af loudnorm=I=-14:TP=-1.5:LRA=11`.
- Fade in/out; no clipping; no abrupt audio cuts; sound design synced to
  cuts is cheap polish that reads as taste.

From maniculehq/creative-duel, MIT.

More from maniculehq/creative-duel