Lyria 3.5 prompts for video BGM: section tags and timestamps

To score a video with Lyria 3.5, put the structure in the prompt: section tags such as [Verse], [Chorus] and [Bridge], and timestamp ranges such as [0:00-0:10]. Google's guide lists both, and Sume's docs show the same timestamp style, because there is no duration parameter to set. The prompt is the only length control you have.

What the controls are

Facts from Google's music generation guide and Sume's Music docs.

The seven-axis brief

Sume's docs recommend a brief with 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. The docs call these creative directions, not guaranteed output values, and tell you to listen to the result. For video, the named moment is where the timestamp earns its place: write it as a time, such as the lift at 0:20 where your product appears.

Contrast matters when you score many scenes in one project. Sume's guidance is to change the broad genre family, move the tempo at least 12 BPM, and change the lead instrument between scenes, unless you want a single consistent score.

A 15-second spot

This prompt asks for an instrumental with three timed sections. The tempo, key and instruments are examples, not recommendations; swap them for your brand.

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: bgm-spot-001" \
  -d '{
    "model": "lyria-3.5",
    "prompt": "A 15-second track. Upbeat, proud, 112 BPM, G major. Rhodes chords, plucked bass, light claps. [0:00-0:05] Intro: Rhodes alone. [0:05-0:11] Verse: bass and claps enter. [0:11-0:15] Chorus: full band, clean ending. Instrumental, no vocals."
  }'

After the call

Poll GET /v1/jobs/:id/status, then read the audio artifact from result.artifacts[] where type is audio. When the provider reports it, result.lyrics holds a lyrics or section map. Sume describes this as model-reported metadata, not a measurement, so open the file and check that the lift really lands where the brief said. Google's guide says outputs vary across identical calls and the model is single-turn, so a miss means a new generation with an adjusted prompt, not an edit.

At Sume's fixed $0.125 per accepted generation, ten attempts at a spot cost $1.25. Keep the prompt that worked next to the final file.