Keep only the sound effects: remove the music, keep the punches

A request shows up regularly in editing communities: I have a fight-scene edit synced to a song. I want to keep the punches, the whooshes, the impacts — and lose the song underneath. Or the tournament montage where the venue music drowns the ball hits. Or the car clip where you only want the engine note.

What these have in common is a specific kind of separation: not "voice versus music," but "music versus everything else." The sound effects, ambience, and foley are the part worth keeping; the licensed song layered on top is the part that has to go — usually because of a copyright claim.

Here's how that job actually works, what the limits are, and the fastest way to get a finished video instead of a puzzle of stems.

Why "keep the SFX" is its own separation job

Standard AI separation splits audio into a vocal stem and an instrumental stem. The instrumental contains drums, bass, keys — and, crucially, any non-musical sound: punches, footsteps, door slams, crowd noise, engine revs.

If your licensed song is playing under the footage, the vocal stem is mostly the singer. The instrumental stem is the band plus your sound effects, mixed together.

So "remove the music, keep the punches" resolves to: keep the instrumental stem, drop the vocal stem. The song's vocals leave; its instruments leave with the music side only insofar as the model can distinguish them from the SFX — which, when the song is dense, is imperfect. The result is usually very usable: your effects remain, the song collapses to a faint instrumental ghost or disappears entirely, depending on the mix.

If instead your goal is the reverse — kill everything except a voice (interview, tutorial) — that's the same pipeline with the opposite choice of which video to keep. One separation run produces both directions; you just pick the output.

What quality to expect

Honesty first, because this workflow has real limits:

• Sparse songs over clean effects separate remarkably well. A beat with distinct hits over crisp foley is the easy case.
• Dense wall-of-sound songs bleed into the effects layer. You may hear a ghost of the chorus under your impacts.
• Effects that resemble instruments — impacts are basically percussion — can partly migrate into the vocal stem. Test before committing.
• The video side is unaffected either way. Separation is an audio operation; your picture is copied, not recomputed.

This is why the "test on 30 seconds first" rule exists in every serious tool: the outcome depends on your specific mix, and no honest provider pretends otherwise.

The fastest workflow

The old route: extract audio in an editor, run a desktop separator, import the instrumental stem back, align it, re-encode. Twenty minutes of timeline surgery per clip, plus one quality-losing export.

The short route with BGM Remover:

1. Sign in — a one-time 5-minute free trial is included, no watermark on outputs.
2. Drop the video. MP4, MOV, MKV, WEBM up to 500 MB and 30 minutes. The video file never leaves your device; only the extracted audio track (a few MB) goes to the AI and is deleted within 24 hours.
3. Download the music-removed video. The picture is bit-identical — the video stream is copied, only the audio is replaced — and the effects that lived in the non-vocal part of the mix come back with it.
4. If you want surgical control afterward, grab the two stems as well and keep working in your editor.

Step 3 is the part that saves the evening: no timeline, no alignment, no render. The deliverable is a finished video with the licensed song stripped out and your own sound design intact — ready to re-upload without the claim.

When this is the right job for separation

• Beat edits and fight choreography synced to a song you can't license — keep the impacts, swap in a track you own.
• Sports and tournament footage with venue music — keep ball hits, crowd, and commentary; lose the arena playlist.
• Car, drone, and gear clips — keep the mechanical sound, remove the music a well-meaning editor added.
• Copyright rescue — the claim was on the song; your effects are innocent bystanders that separation lets you keep.

Frequently asked questions

Can I remove the music but keep the dialogue AND the effects?

Yes — both dialogue and effects are non-vocal-layer content, so they travel together in the instrumental side. This is the standard "copyright rescue" recipe: claimed song out, voices and foley stay.

The instrumental ghost of the song is still faintly audible. Can I clean it further?

Sometimes. Running the resulting track through a second pass can reduce residue, at some quality cost. Often a better fix: duck the remaining ghost under your replacement music — a well-chosen new track masks small artifacts completely.

Does this work on audio files too, or only video?

Both. Video input returns the remuxed videos; audio input returns the two stems. The separation is the same either way.

Why not just lower the music volume instead?

If the goal is beating a Content ID fingerprint, volume reduction doesn't work — fingerprinting survives level changes. If the goal is just listenability, volume is fine. Separation is for when the track has to actually be gone.

Try it on your own edit

Drop the clip, sign in, and one pass returns the effects-complete video with the song gone — remove the music, keep your sound design.