Removing music from a video is either a ten second job or a genuinely hard one, and which you get depends on something you can check before you install anything. If the music sits on its own track, mute it and you are done. If the music was recorded into the same mix as the voice, no slider takes it out. A model has to guess which part of the sound is the voice, and it will guess wrong about some of it.
This page covers both cases, in the order to try them. The first case is free and perfect. The second is neither.
First, find out which problem you have
| What you have | What works | What it costs you |
|---|---|---|
| Music on its own track | mute or delete that track | nothing, the voice is untouched |
| Music mixed with the voice in one file | source separation | marks on the voice |
Open the video in an editor and look at the timeline. Audio tracks are listed separately, so a clip with a voice track and a music track under it is the easy case, even if the two are mixed down in the file you exported last time.
A file you downloaded, a screen recording with a song playing in the room, or a phone video from a party is the hard case. One track, everything inside it.
1. If the music has its own track, mute it
Every editor puts a mute control on the track header. Press it and play the timeline. The music is gone and the voice is exactly as it was, because nothing was processed.
If you would rather not keep an empty track, select the music clip and delete it, then close the gap so the rest of the timeline does not slide.
Then export. The file you get has silence where the music was, and the voice has lost nothing.
2. If the music is mixed in, the job is source separation
Separation is a different kind of tool from a filter. A noise remover looks for a steady hiss and turns it down. A separator listens to the whole mix and estimates which parts belong to the voice, which to drums, which to bass, and which to everything else, then writes each estimate to its own file.
That is why the result is never clean. The model is guessing, and where the music and the voice overlap, it has to choose. Check whether your editor has a voice isolation or stem separation control before installing anything. If it does, try it on a short section first, since these features vary a lot between versions.
3. Separate the voice with Demucs
Demucs is the free open source route. It is MIT licensed, it comes from Meta’s research group, and version 4 is the current release.
Install it with Python:
python3 -m pip install -U demucsIt pulls in PyTorch, which is the large part of the download. It runs on a CPU, and a GPU with about 3 GB of memory makes it faster.
Give it an audio file. If what you have is a video, take the sound out first. Gleam Audio opens mp4, mov, mkv, webm and avi for their sound since 2.0 and saves the result as m4a, and your editor can export an audio track too. The two way round, putting audio back into a video, is covered in how to remove audio from a video.
Then run it:
demucs --two-stems=vocals clip.wavThe two files land in separated/htdemucs/clip/: vocals.wav and no_vocals.wav, both stereo at 44.1 kHz. Leave the flag off and you get four stems instead, drums.wav, bass.wav, vocals.wav and other.wav.
Two things from the project’s own README are worth knowing before a long job:
--two-stems=vocalsseparates the mix in full and then adds the other stems back together, so it is not faster and does not use less memory than the four stem run.-n htdemucs_ftis a fine tuned model that takes about four times as long and may be slightly better. Use it when the first result is not good enough.
The repository is no longer maintained by Meta. The author keeps a fork at adefossez/demucs for important fixes.
4. Clean the vocal stem before it goes back
The vocals file is not a clean voice. It has music bleeding through, and separation leaves a thin, slightly hollow edge on the speech. A pass with a speech tool fixes part of that.
Remove noise in Gleam Audio works on the stem, offline, with nothing uploaded. It is free for personal use up to 20 files a week, and the license page lists what each tier allows.
Keep the strength low. The model behind Remove noise is trained on speech in real rooms, not on separated stems, so a high setting on a file that has already been through a separator tends to eat the consonants. Start at the bottom of the range and use the Before and After switch on the finished file to hear the difference while it plays.
Do not run it on the music stem if you are keeping the music. The known limitations page says the tool is tuned for speech, and on music it can take out more than you want.
5. Put the stem back on the timeline
Import vocals.wav and drop it on an audio track lined up with the start of the original clip. Mute the original clip’s audio rather than deleting it, so you can switch between the two while you work.
Sync holds. Separation keeps the length of the file, so the stem starts and ends where the original audio did, and a lip sync check on the first sentence will confirm it.
One thing to watch: Demucs rescales each stem so it does not clip, and the README warns that this can change the relative volume between stems. If the voice now sits louder or quieter than it did in the mix, ride the clip gain rather than the track fader, so the rest of the timeline keeps its balance.
6. Listen for what separation broke
Put headphones on and listen to a paragraph, not a phrase. The faults are consistent:
- Consonants. S, t and k are the first sounds to smear, and a voice that has lost them sounds like it is speaking through cloth.
- Sibilance. The model can leave a lisp, or an s that is too bright against the rest of the word.
- Room. Reverb tails get pulled apart. The voice can end up sounding like it was recorded in a smaller room than it was.
- Level between words. Music bleed that used to sit under the speech is now a small swell in every pause.
If the result fails two of those, go back and try the fine tuned model, or accept the mix and keep the music.
When it won’t help
- Music louder than the voice. There is more music than voice in the mix, so the model has less to work with and the voice comes back thin.
- Singing. The vocals stem keeps the singing, so a separator cannot pull a singer out of a band and leave you the band, or the other way round.
- A phone video with a song playing in the room. The music arrives through the same microphone as the voice, with the room on top of both.
- Live recordings with reverb. The tail of the room is in both signals, and separation spreads it across the stems.
- A heavily compressed source. A low bitrate mp3 leaves the model less detail to work from, and it tends to bring the codec artefacts up with the voice.
Takeaways
- Check the tracks before you install anything. If the music is separate, muting it is the whole job.
- If the music is mixed in, only source separation helps, and it always leaves marks on the voice.
- Demucs is the free route: install with pip, then
demucs --two-stems=vocalson the audio file. - Clean the vocal stem at a low strength, put it back on the timeline, and listen to the consonants before you export.