Kids video background music: a Lyria prompt that stays instrumental

The prompt

Describe the music in one paragraph that names emotion, genre, tempo, key, instruments, arc and era, state the length in words, and end with "Instrumental, no vocals." One Sume generation costs a fixed $0.125 and returns one track you can lay under narration.

Example: "A warm, playful 2-minute track for a children's counting video. Light folk-pop, 100 BPM, C major. Ukulele, soft glockenspiel, upright bass and brushed drums. Starts sparse, builds gently at 0:45, resolves softly. Late 2010s acoustic feel. Instrumental, no vocals."

Why each axis is there

Sume's music router takes a free-text prompt of 1 to 5,000 characters. It has no duration field, so a duration value is rejected; put the length in the prompt instead. The music router docs describe the model choices (sume/music-auto by default, lyria-3.5 or lyria-3-pro).

What not to put in the prompt

A non-empty negative_prompt returns a 400 with negative_prompt_unsupported, so say what you want rather than what to avoid. The closing line "Instrumental, no vocals." does the work a negative prompt would do elsewhere.

Do not name a living artist or paste someone's lyrics. Google's Lyria documentation (read 2026-10-04) says it blocks artist-voice and copyrighted-lyric prompts, and Sume routes to Lyria, so describe a style in your own words.

Three variations worth trying

Vary one axis at a time so you can hear what changed. Keep the tempo and key and swap the instruments for marimba and pizzicato strings for a lighter, quirkier bed. Or keep the instruments and slow the tempo to 80 BPM for a calm bedtime or story segment. A third version can drop the arc, asking for a steady, loop-friendly pattern, if the video is long and the music should not draw attention.

Because every generation is a fixed price, three variations cost $0.375 in total (3 x $0.125). That is a cheap way to choose, and far cheaper than re-cutting a finished video around the wrong track.

Get it into the video

Submit POST /v1/music-router/generate with a prompt and an Idempotency-Key. Poll the job, then read the track from result.artifacts[]. The job's request.routed_model shows which engine ran. To lay it under a narration, mix it in your editor or Sume's timeline step, and keep the bed quieter than the voice.

• Make one track first; the generation is not a place to iterate blindly at $0.125 each.
• Listen for any vocal-like sounds before you publish.
• Check the platform's rules for children's content separately; the music prompt does not settle them.