AI sonic logo generator: a short audio sting by prompt, then trim

To make a sonic logo with AI on Sume, ask the Music Router for a very short sting in the prompt, then cut the file to length with Timeline audio. There is no duration field on the music API, and sending one is rejected, so length is a request in words and a trim in a second job. Expect to run it a few times, because a sting is judged in the first second and a miss is obvious.

A sonic logo is a harder ask than a background bed. A bed can drift. A sting has to land: one gesture, one timbre, a clear resolution, and an ending that does not trail. This post covers a prompt shaped for that, the trim step, what Google's page says about how short a Lyria 3.5 track can be, and what Sume does not offer.

How short can the track be?

The Music docs say to steer length in the prompt with phrasing such as "a 2-minute track". The same works at the other end: "a 3-second sting". Google's Lyria page describes the short end of a Lyria 3.5 track as a 60-second clip, so treat a request for a few seconds as a hint the engine may not honor. Put the sting at the very start of the track and plan to cut.

What should a sonic-logo prompt say?

Write the sting as a single event. Name the timbre and the shape of the gesture, say where it resolves, and say that nothing follows. Avoid tempo talk; a sting has no beat to keep. Close with the same clause the docs recommend, so no vocals or spoken word appear.

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: sonic-logo-001" \
  -d '{
    "prompt": "A 3-second sonic logo. A soft glass-bell note rises a perfect fifth and lands on a warm major chord with a short marimba pluck underneath, then a natural decay to silence. One gesture, clean and confident, no melody after it. Instrumental, no vocals, no spoken word."
  }'

How do I cut the file to the sting?

Once the job completes, take the artifact URL and cut it. Timeline audio split slices one Sume-hosted file into up to 20 ranges, each with its own durable audio_url. The default output is wav. The job is a flat $0.01, with no provider inference, and it runs on the worker's ffmpeg, as the Timeline audio page notes.

curl -X POST https://api.sume.com/v1/timeline-1.0/audio \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: sonic-logo-cut-001" \
  -d '{
    "operation": "split",
    "url": "https://media.sume.com/artifacts/artf_demo/sting.mp3",
    "ranges": [{"start": 0, "end": 3.2}, {"start": 0, "end": 2.4}],
    "output": {"format": "wav"}
  }'

How do I audition the cut?

Two ranges in one call gives you a longer and a shorter version without a second job, since ranges may overlap. Listen to both on a phone speaker as well as headphones, because a logo is often heard on a laptop or a TV. If the cut chops the decay, move end later; if it leaves silence, move it earlier.

A logo also lives in videos. You can pass the cut file to a Timeline render as the soundtrack of an end card, and the render's own fade_out_seconds on the soundtrack reaches up to 10 seconds, so a sting can sit at the tail of a clip.

Should I keep regenerating the sting?

Once you pick a sting, treat the file as the asset and stop regenerating. The Music docs list no seed, temperature or guidance parameter, so the same prompt will not return the same sound, and a brand sound that changes on every render is not a brand sound. Download the chosen wav, store it with your brand files, and refer to that copy from then on. Sume's media.sume.com artifact URL is the durable copy on its side, but your own archive is the one you control.

Do the same for the longer and shorter variants you cut. Name them for where they are used, such as end card, podcast open and ad tail, so nobody has to guess later.

What does Sume not do for a sonic logo?

Sume does not guarantee that a prompt yields a given length, tempo or note. It does not make trademark or sound-mark filings, and nothing in its docs says a generated sting is clear of other marks. For a logo that stands for a company, keep the prompt, the date, the engine named in job.request.routed_model, and the cut file together, and take your own advice before you register or rely on it.

If you want a mix of a voice tag and a sting, that is two separate jobs on Sume: a text-to-speech job for the voice and a music job for the sting, joined with Timeline audio concat.