Lyria 3.5 section timestamps: write a 30-second ad bed prompt

Google's Lyria 3.5 docs show section timestamps such as [0:00 - 0:10] inside the prompt (read 2026-10-03). Sume's Music Router has no duration field, and its docs steer length in the prompt with the same idea, for example [0:00-0:30] Intro: .... A three-part, 30-second bed fits in one prompt.

A three-section brief

Give each block a start, an end, a role and a sound. Keep the total at 30 seconds and say the track is instrumental if you plan to put speech on top.

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: ad-bed-001" \
  -d '{
    "model": "sume/music-auto",
    "prompt": "A 30-second instrumental ad bed, 100 BPM. [0:00-0:05] Hook: soft pluck and a riser. [0:05-0:22] Body: steady beat, warm bass, space for a voiceover. [0:22-0:30] Outro: resolve and stop on a clean final chord. No vocals."
  }'

What to expect back

The result has an audio artifact in result.artifacts[] and result.lyrics, which carries the model-reported section map. Compare the map with your brief: the model reports what it made, which can differ from what you asked for. Output is non-deterministic, so generate a few takes.

Fitting it to the spot

If the track runs long, the render can cut it with a fade_out_seconds of up to 10. If it runs short, set loop on the soundtrack. Put the voice on the render's audio.url and use duck_db so the bed steps back. Compare engines with the catalog endpoint post.