Launch teaser music on Lyria 3.5: section markers, 30 seconds

To score a 30-second launch teaser, send a prompt to POST /v1/music-router/generate that states the length and uses section markers like [0:00-0:10] Intro: .... There is no duration field, so length is steered by the prompt and verified by listening; each generation costs a fixed $0.125 on Music 1.0.

Which model does the Music Router use?

Omit model or send sume/music-auto and Sume picks the engine, Lyria 3.5 today. Explicit ids lyria-3.5 and lyria-3-pro pass through. job.model echoes the id you asked for, and job.request.routed_model names the engine that ran. Music 1.0 is retiring gradually and now resolves through the router, so use the router for new work.

What does a teaser prompt contain?

The docs suggest a scene-specific brief across seven axes: emotion, genre, tempo as a number, key and mode, two to four instruments with texture, an arc with one named moment, and era or production. These are creative directions, not guaranteed settings. For a teaser the arc matters most, because a reveal needs a lift.

What does the call look like?

Prompts are 1 to 5000 characters. Put exclusions in the positive prompt, because a non-empty negative_prompt is unsupported. Closing with "Instrumental, no vocals" keeps the track clear for a voiceover line.

curl -X POST https://api.sume.com/v1/music-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: teaser-score-001" \
  -d '{
    "model": "sume/music-auto",
    "prompt": "A 30-second launch teaser track, synthwave, 112 BPM, A minor. [0:00-0:10] Sparse analog bass pulse. [0:10-0:20] Muted pluck and tight claps build. Full drop at 0:20 for the reveal. Instrumental, no vocals."
  }'

How do you land the track on the cut?

Poll GET /v1/jobs/:id/result and read the audio artifact where type is audio. Length is steered by the prompt only, so check it. Import the track, then use it as the Timeline 1.0 soundtrack with loop, fade_out_seconds (up to 10) and duck_db if there is a voiceover. Ducking needs a real spine, not silence.

If the reveal lands at 0:21 instead of 0:20, nudge your cut rather than regenerating at once: shift the reveal slot's start to the drop. Regenerate only when the arc is wrong.

How do you iterate without overspending?

At $0.125 per generation, a handful of takes is cheap compared with video generation, but it adds up across a campaign. Change one axis at a time, for example tempo or the lead instrument, so you learn what each change does. Keep a note of the prompt that worked, since there is no seed to reproduce a take.

For several teaser versions, vary broad genre family, tempo by at least 12 BPM and lead instrument, so they do not sound like one track. For a consistent campaign score, keep those three the same and change only the arc.

What should you check in the finished track?

Listen for four things: the length against your cut, the point where the drop lands, whether any vocals slipped in despite the instruction, and how the opening sits under a voiceover line. The docs describe the brief as creative direction and tell you to verify the generated audio, so treat every take as a draft until you have heard it against the picture.

Then import it and set the Timeline 1.0 soundtrack gain and fade. Keep fade_out_seconds within its 10-second limit.

What does this not do?

Music has no seed parameter, so do not expect to reproduce a take from the same prompt. Image conditioning exists through image_url, but results are not guaranteed to match the picture. Check the audio and the licence terms that apply to your use before you ship.