Sound & Music in H3 (Without the Jargon)
H3 generates real audio in the file — here is how to split ambience, foley, dialogue, and score so it doesn’t turn to mush.
A silent-visual prompt still gets random audio. H3’s differentiator is native stereo sound — treat it as half the job.
Two buckets (never merge them)
| Bucket | Official field | What belongs here |
|--------|----------------|-------------------|
| Sounds in the world | overall_soundscape | Rain, crowd bed, footsteps, dialogue, engine, UI clicks |
| Score for the viewer | non_diegetic_music | Piano, drums, synth pad — or N/A |
Putting “sad piano under rain” only in the soundscape line confuses the model. Split them.
Dialogue = performance, not caption
Weak: “She talks about the product.”
Strong: “(S1) turns to camera, quiet intensity: ‘It was never about the money.’ Room tone continues under voice.”
Keep lines short — count seconds on your fingers.
Music that edits well
For ads and title sequences, tie cuts to beats:
Hard cut at 3s and 6s on kick; riser into final logo; silence last 2 seconds before end card.
Silence at the end is a pro move — it makes logo stings land.
Levels of detail (pick one)
Light — “Coffee shop ambience, soft jazz low in mix.”
Medium — “Espresso machine hiss, cup clink at 4s, distant chatter −20 LUFS feel.”
Heavy — “0–3s: only kettle; 3–6s: pour crescendo; 6s+: add brushed drums at 80 BPM.”
Heavy helps music-driven spots; light is fine for talking-head style.
Common audio mistakes (from docs + tests)
- Writing “cinematic audio” with zero nouns
- Scoring and ambience in one sentence
- Long monologue in a 10s clip
- Forgetting to say when music stops
Practice path
- Read How prompts work
- Open any commercial template — scroll to How this prompt is built
- Change only the sound lines first; keep visuals fixed — learn cause and effect
Next: Common mistakes