AI music adds vocals under narration: the Sume prompt fix

If AI music keeps adding vocals under your narration, do not send a negative_prompt to Sume: a non-empty value returns HTTP 400 with public_reason=negative_prompt_unsupported. Put the exclusion in the positive prompt instead. End with "Instrumental, no vocals." and, only when the track sits under a voiceover, add "no spoken word." Each attempt is a $0.125 generation, so three tries cost $0.375.

What the docs say to write

The Music 1.0 docs list the request fields. prompt is 1 to 5,000 characters, and exclusions belong inside it. negative_prompt is accepted only as an omitted field or an empty string. The same docs advise ending a brief with one clause, "Instrumental, no vocals." and adding "no spoken word" only under narration, because a bed that contains speech competes with the voiceover.

A bed brief that keeps out of the voice

Write for the space the voice needs. Name two to four sparse instruments with texture, give a tempo as a number and a key and mode, and describe the arc: for a bed, that is usually flat. A bed with a clear melody in the 1 to 3 kHz range fights speech, so choose pads, low bass and a soft pulse. Length is in the prompt, such as "A 45-second track", because Music 1.0 does not take duration or duration_seconds.

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-under-vo-01" \
  -d '{
    "prompt": "Warm ambient bed, 70 BPM, A minor. Soft analog pad, low sine bass, a faint tape-hiss pulse, no melody. Flat arc, no build. A 45-second track. Instrumental, no vocals, no spoken word."
  }'

Why the vocal clause is not a guarantee

A prompt is guidance for a generative model, not a filter. Sume documents the exclusion as a recommended clause, not as a switch, so a track can still contain a vocal-like texture, a hummed pad or a breath. That is why the order matters: place the clause last, where it reads as the final instruction, and keep the rest of the brief free of words that suggest a singer.

The cost of testing is small and fixed. At $0.125 per generation, five variants of one brief are $0.625. Generate them together with different Idempotency-Key values, listen to all five, and keep the one that stays out of the voice's way. Reusing the same key would return the first job, not a new take.

Choosing the router or the direct route

The Music Router takes the same prompt rules and picks an engine for you; the direct Music 1.0 surface is the fixed one. Both bill $0.125 per generation. After a job finishes, check which engine actually ran before you compare takes, since a routed job can differ from the requested one. The track length still comes from the words in your prompt, so say "A 45-second track" every time and compare lengths in the result.

If vocals still appear

Four steps, in the order that costs least.

• Listen first. Provider lyrics, when present, are model-reported metadata, not a measurement of what you hear.
• Remove words that invite singing: "anthem", "choir", "soulful" and any lyric-like phrase in quotes.
• Change one thing per retry, and write down the job id, so you know which edit worked.
• Lower the bed under the voice anyway. In a Timeline render soundtrack.duck_db (0 to 20) lowers it under speech, and a bit of leaked vocal is easier to hide at -12 dB.