Gemini Omni clip has sound I did not want: how to remove the audio

On Sume, Gemini Omni cannot be switched to silent output, so there are two steps: prompt away what you can, then remove the track. Google's Omni guide says the model by default tries to generate an appropriate audio track, and Sume's Video Router doc says native audio is always on for gemini-omni-flash-1.1. Do not count on a generate_audio: false flag to mute it. Instead, prompt for no dialogue and no extra sound effects, then drop the audio with a Sume video trim.

What can the prompt do about unwanted audio?

The guide has a "Removing unwanted elements" list with short phrases: "No dialogue", "No embellishments" and "No extra sound effects". It also says negative prompts are not supported as a field, and that you can put negatives in the regular prompt, for example "Do not do X". For music, the guide's examples are to include calm background music or a high-energy techno beat, which steers the track rather than removing it.

How do I drop the audio track?

Sume's video trim takes one clip plus a range and returns a new MP4. Its audio option is keep (the default) or drop, so trimming the whole clip with audio: drop returns a silent copy for $0.02 per job. The source stays untouched. Set start to 0 and duration to the clip length; the doc says a range past the source clamps and warns, so a slightly long duration is harmless.

The trim source must be a media.sume.com artifact or asset in your workspace. The trim doc says there is no open-internet fetch, so import a clip first with POST /v1/media-imports if it is not there already.

curl -X POST https://api.sume.com/v1/video-trim \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: omni-silent-001" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_demo/omni-clip.mp4",
    "start": 0,
    "duration": 10,
    "audio": "drop"
  }'

Does a reference clip carry its sound into the result?

Not according to Google. The Omni limitations list says video references work best with likenesses and that any audio in a video reference is ignored, and it says uploading audio references is unsupported in the current version. Sume's catalog matches: gemini-omni-flash-1.1 has no reference_audio_urls. So the soundtrack you hear comes from the prompt and the model's own choices, never from a clip you pass in. If the voice or music keeps appearing, the fix is in the prompt wording, not in the references you attach.

For edits the same logic applies. A clip sent as video_url is an edit source, not a reference, so check probe.has_audio on an edited output before assuming it matches the source, and run the same drop step if you need silence.

How do I confirm the clip is really silent?

Run Sume's video inspect on the trimmed result with frames: false, which returns probe facts only, and read probe.has_audio. The audio detach doc names the same check: it says a source with no audio track fails with detach_source_has_no_audio, and recommends checking probe.has_audio first. A silent trim should report no audio.

Silence is often a step toward a better soundtrack. Once the track is gone, you can lay your own music under the clip with Sume's Timeline, whose soundtrack field takes a bed with gain and fade options. If you would rather keep the native sound and add music on top, extract it with audio detach and mix it down instead.

Keep the original clip. Trim creates a new file and leaves the source alone, so you can always go back to the version with sound if the silent cut turns out to be flat.