AI ambient rain or white noise track: a Lyria prompt and a loop

You can make an ambient bed such as rain, wind or a soft drone with Sume's Music Router, but you are generating a short piece of music, not a seamless sound-effect loop, so loop it in the render rather than expecting the file to repeat cleanly. Prompt for texture and an even arc, say instrumental, then set loop: true on the Timeline soundtrack.

Sume's route is documented in the Music Router guide, and the model behavior is from Google's music generation docs, read on 2026-10-02.

Does Sume have a sound-effect or noise endpoint?

No. The public API has no sound-effect or noise endpoint; its audio routes are text to speech, speech to text, music and file tools such as audio detach and timeline audio. Music runs on Lyria 3.5, which Google describes as generating full compositions of a couple of minutes in MP3 or WAV, controllable using the prompt. A rain recording or a pink-noise generator is not what it is built for, so treat the result as an ambient piece, listen for a pattern you do not want, and choose a take that stays even.

Also note what the route cannot do: no duration field, no negative prompt, and no seed, so identical prompts give different takes.

How do you prompt for an ambient bed?

Describe the texture and an arc that does not build. Add the length you want in words, because there is no duration parameter, and put exclusions in the positive prompt. Give two or three instruments or sound sources, a tempo or an explicit lack of rhythm, and close with the instrumental line. Section markers like [0:00-0:30] let you hold one mood across the whole track.

Here is a request using the router:

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: ambient-001" \
  -d '{
    "model": "sume/music-auto",
    "prompt": "A 90-second ambient bed for focus. Soft evening rain texture, a slow warm synth pad in D minor, a distant low drone, no drums and no melody. Even from start to end, no build, no drop. Instrumental, no vocals, no spoken word."
  }'

How do you loop it under a longer video?

Take the audio URL from result.artifacts[] (typically audio/mpeg on media.sume.com) and use it as the Timeline soundtrack. The soundtrack object takes url, gain_db, loop, fade_out_seconds up to 10, and duck_db from 0 to 20. When a bed is shorter than the video, Timeline can pad or loop it with a warning, and loop: true makes that explicit.

A loop point is the weak spot: the end of the track meets the start. Pick a take that ends as it begins, and use a short fade at the end instead of a hard cut. If you hear a seam, ask for another take, because the same prompt will give a different one.

What should you check by ear?

Listen for a repeating phrase, a drum that was not requested, or a vocal vowel in a track marked instrumental. Any of these means you pick another take. A fixed price of $0.125 per generation makes it cheap to try several, and Music Router bills the same on every routable model.

For a long video, one short bed repeated for an hour will wear on a viewer. Generate two or three different beds, join them with timeline audio concat into one file, and use that file as the soundtrack. Keep the level low under speech, using gain_db and duck_db, so the bed supports the voiceover instead of competing with it.

Keep the job id, the prompt and the model with the final file. If the audio is part of an ad, check the label rules for AI-made music on the platform where it will run.