A 90-second music bed under a voiceover: one prompt, 12.5 cents

To get a 90-second bed under a voiceover from Sume's Music Router, write "a 90-second track" and timed section markers into the prompt, end with "Instrumental, no vocals" and add "no spoken word" because narration will sit on top. The job costs a fixed $0.125 whatever the length or the engine you pin.

Why the length lives in the prompt

The Music Router and Music 1.0 pages both say Sume rejects duration and duration_seconds. The prompt controls length: the docs give "a 2-minute track" and markers such as [0:00-0:30] Intro: ... as examples. Music 1.0 runs on Google Lyria 3.5 and makes structured songs of up to a few minutes, so 90 seconds is inside the range the docs describe.

The brief is a direction, not a guarantee. The Music 1.0 page says the axes are creative directions, not guaranteed output values, and tells you to examine the audio you get.

A prompt with three timed sections

Voice-led ads have three beats: a hook, the proof and a call to action. Give the bed an arc that leaves space under the voice in the first two and lifts at the end. Put the exclusions in the positive prompt, because a non-empty negative_prompt returns a 400 with negative_prompt_unsupported.

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: bed-90s-001" \
  -d '{
    "model": "sume/music-auto",
    "prompt": "A 90-second track, warm indie-pop, 96 BPM, G major. [0:00-0:30] Hook: soft Rhodes and a muted kick, sparse. [0:30-1:00] Proof: add brushed snare and a low bass, still quiet in the mids. [1:00-1:30] Call to action: open up with claps and a bright lead. Instrumental, no vocals, no spoken word."
  }'

Auto or pinned

If you omit model or send sume/music-auto, Sume picks the engine (Lyria 3.5 today). To pass through to one engine, pin a catalog id from GET /v1/music-router/models. The routable ids are sume/music-auto, lyria-3.5 and lyria-3-pro. The job echoes the id you requested, and job.request.routed_model names the engine that ran.

Fitting the bed to the video

The bed will not be exactly 90 seconds, so read the result's audio duration and decide before you render. A Timeline 1.0 render takes audio.duration_seconds from 1 to 1,800 and bills $0.10 per ceil output minute, so a 90-second render is billed as 2 minutes, or $0.20. If your bed is 84 seconds and the voice is 90, trim the voice slot or re-roll the bed; a second bed is another 12.5 cents.

The timeline carries one audio spine. To put the voice and the bed on one track, mix them outside Sume or choose which one the spine is. The Timeline audio page covers joining parts, not mixing, so do not assume a ducking control.

Budget the retakes

Plan on two or three beds for one ad, since the direction is not a guarantee: 25 to 37.5 cents. A policy rejection means you change the flagged content and keep the musical brief rather than reduce it to a generic bed.

Checklist before you queue

Keep the prompt in your repo next to the script. When the voice changes length, you edit the markers and generate again, and the file shows exactly what you asked for last time.

• Length: say the number of seconds once, near the start of the prompt.
• Structure: three timed sections, each with a one-line instrument note.
• Space for the voice: ask for sparse mids early.
• Exclusions: end with the instrumental and no-spoken-word clauses.
• Key and tempo: name them if the bed must sit with a jingle.

What a rejected brief looks like

Because the router refuses a non-empty negative_prompt and any duration field, the two commonest mistakes fail fast with a 400 and no charge for a job that never started. Fix them by moving the exclusions into the prompt and the length into a sentence. A rejected brief is the cheapest kind of retake, which is a reason to send a first request from a script and read the error text instead of from a guess.