How to Remove Music From a Video Without Wrecking the Speech
There are two different jobs here. If nothing else in the audio matters, mute the video's sound track in any editor — Photos on iPhone, Google Photos on Android, CapCut or DaVinci Resolve on desktop — and export; this is quick and lossless in quality terms because you are discarding a stream, not reprocessing one. If someone is speaking over the music and you need the speech to survive, you need source separation, which uses a trained model to split a mix into parts. It works well on music mixes and imperfectly on speech-over-music, and it can never recover detail that the mix destroyed.
The short answer: There are two different jobs here. If nothing else in the audio matters, mute the video's sound track in any editor — Photos on iPhone, Google Photos on Android, CapCut or DaVinci Resolve on desktop — and export; this is quick and lossless in quality terms because you are discarding a stream, not reprocessing one. If someone is speaking over the music and you need the speech to survive, you need source separation, which uses a trained model to split a mix into parts. It works well on music mixes and imperfectly on speech-over-music, and it can never recover detail that the mix destroyed.
Work out which question you are asking
Almost every frustrating hour spent on this problem comes from starting with the wrong one of these.
Question one: is anything else in this audio worth keeping? If the answer is no — a wedding clip with a radio in the background, a phone video with a shop playlist over it, a piece of footage you intend to score yourself — you do not have an audio problem. You have a mute button.
Question two: do I need the speech to survive? If someone is talking, singing or narrating and you need that, you are in a completely different technical territory, one where the honest answer is a set of trade-offs, and no setting avoids them.
The rest of this guide is those two answers in order — and then the third option, which is not to inherit the problem in the first place. If you are stripping a track because you intend to put your own under it, how to make an AI music video is the version of this job that starts from the song.
Muting, which is easy and worth doing right
On iPhone. Open the video in Photos, tap Edit, and use the audio control to turn the clip's sound off. Save.
On Android. Google Photos will mute a video's own audio outright while you edit it — on most phones that is a single Mute tap. In a Photos creation with a soundtrack over it, separate volume sliders for "Video audio" and "Soundtrack" let you balance the two instead of choosing between them.
On desktop. In CapCut, DaVinci Resolve, iMovie or Premiere, detach or select the audio track and set its level to silence — or delete the track outright, which is cleaner.
Two things worth getting right while you are there.
Silence is not always the goal. A completely silent video reads as broken to a viewer, especially on a platform where sound is expected. Replacing the music with room tone, or with something you are licensed to use, is usually a better outcome than removing it and shipping nothing. The routes for adding a new track are the same ones in reverse.
Export once. Muting itself costs no quality — you are dropping a stream, not reprocessing it — but every full export re-encodes the picture. Do the whole edit, then export.
When a copyright claim is the reason you are removing it
A special case with its own tooling. If the reason you want the music gone is a copyright claim on YouTube, you do not need to re-edit and reupload at all.
YouTube provides a way to remove copyright-claimed content from a video in place, and the option that matters is muting: you can mute only the claimed song and leave the rest of the audio in place, and if all the claimed audio can be muted, the claim is removed from the video. That matters because a Content ID claim brings restrictions with it — the copyright owner's policy decides whether your video is blocked, monetised for them, or merely tracked — and clearing the claim clears the restrictions.
Check what the mute actually costs you before you accept it. If the claimed track ran under your entire video, muting the song will often take your voiceover with it, at which point you are back to the hard problem below.
Source separation: what it is and what it costs
Source separation models are trained on large collections of music where the individual parts were available separately. They learn what a mix of a given set of parts tends to look like, and can then work backwards from a mix they have never heard to an estimate of each part.
Demucs is the reference point most tools are built on or benchmarked against. It is open source under the MIT licence, and it splits a music mix into four sources — drums, bass, vocals and everything else — with an experimental six-source version that also isolates guitar and piano.
Read that description carefully, because it defines the limit people run into. Demucs separates a music mix into instruments. It is not built to pull a music bed out from under a person talking in a room. Those are related tasks with different training data, and the tools that do the second one well are speech-isolation tools, not stem splitters. Reaching for a stem splitter to clean up dialogue is the single most common way people conclude that "AI audio separation doesn't work", when what actually happened was the wrong tool.
The other limit is harder and does not go away with a better model. Separation cannot recover information the mix destroyed. Where a crash cymbal and a hard consonant occupy the same instant at the same frequencies, the mixed signal contains one number where there used to be two, and nothing can tell you which part of it belonged to the voice. A model produces a plausible guess. The louder the music sits relative to the thing you want to keep, the more of the output is guess rather than recovery — which is why separation on a quiet background bed sounds transparent and separation on a loud club recording sounds like a phone call underwater.
Choosing between them
| Situation | Use |
|---|---|
| Music is the only audio, and you don't want it | Mute |
| You want to replace the music with your own | Mute, then add the new track |
| A Content ID claim on an otherwise fine video | YouTube's in-place mute of the claimed song |
| Speech over a quiet music bed | Speech isolation — good results likely |
| Speech over a loud music bed | Speech isolation — manage expectations, or reshoot |
| You want the instrumental of a song | Stem separation like Demucs — this is what it is for |
The version of this problem worth avoiding entirely
Every route above is remedial. You are undoing a decision that got baked into a file.
If the recurring reason you are here is that a track you used got claimed, the fix is upstream: build the video around music you control. That is not a platitude — it is the specific reason a musician's own release is the easiest video on the internet to publish, and why we build videos from the artist's own track instead of laying a licensed bed over stock footage. If you have a song and no footage, making a music video without filming is the route in, and whether AI music videos are allowed on YouTube and TikTok covers what the platforms make of the result.
Frequently asked questions
How do I remove music from a video without removing the voices?
You need source separation, not muting. A separation model has been trained to split a mixed recording into parts and can pull a music bed down while leaving speech mostly intact. Expect artefacts: thin, slightly watery-sounding speech where the music was loudest, and residual traces of anything that overlapped the voice in both pitch and time.
Does muting a video reduce its quality?
The picture is unaffected if your editor copies the video stream instead of re-encoding it, which most do when only the audio changes. Even when it does re-encode, one export at a sensible bitrate is not visibly damaging. The audio is simply gone, not degraded — you are removing a stream, not processing it.
What can source separation actually not do?
It cannot recover information the mix threw away. Where a cymbal and a consonant occupy the same instant and the same frequencies, no model can tell you what the consonant sounded like on its own — it can only make a trained guess. The louder the music sits relative to the voice, the more of the result is guesswork and the less of it is recovery.
What tools separate music from a video's audio?
Demucs is the best-known open-source model, released under the MIT licence, and splits a music mix into drums, bass, vocals and other, with an experimental version adding guitar and piano. Note what that is doing: separating a music mix into instruments. Pulling a music bed out from under dialogue is a related but distinct task, and dedicated speech-isolation tools handle it better.
How do I remove copyright-claimed music from a YouTube video?
YouTube lets you edit out claimed content without reuploading. From YouTube Studio you can mute just the claimed song and leave the rest of the video's audio alone, and if every claimed segment can be muted, the Content ID claim is removed from the video along with the restrictions attached to it.
Can I remove music from a video on my phone?
Muting, yes — every phone editor has a volume control on the video's own audio track, and pulling it to zero is the whole operation. Separation, not really: the models are heavy, so phone apps offering it are almost always uploading your file to a server and sending the result back. That is fine, and worth knowing before you send footage somewhere.
Stripping the wrong music is a fix. Owning the right one is a plan.
Melodious builds the video around your own track — storyboarded scene by scene from the song you actually control, so nothing in it can be claimed or muted.
Storyboard your song