To remove vocals from a song for free, use an AI separator such as Ultimate Vocal Remover or Demucs, which needs about 3 GB of GPU memory just to start. Audacity’s old built-in effect works without AI, but it stopped shipping with the app in version 3.5.0. This post is for anyone who wants a karaoke track, a practice backing or an instrumental under a video. It covers what each free tool outputs. It also covers what each one needs from your computer, and where it falls apart.
Every vocal remover promises the same thing: the singer leaves. Almost none of them mention who leaves with her.


Free tools that remove vocals from a song, and what each one hands back
Free tools that remove vocals from a song come in two kinds, and that split decides almost everything. One kind cancels whatever sits in the middle of a stereo mix. The other runs a trained AI model that splits the song into stems. Stems are separate audio files for vocals, drums, bass and everything else.
Phase cancelling is instant. It needs no graphics card and works on a ten-year-old laptop. It’s also dumb in a very specific way, which gets its own section below. AI separation is slower and hungrier, but it actually hears the difference between a voice and a bass guitar.
Neither kind can recover what was never recorded separately. A finished song is one mixed signal, and every tool here is making an educated guess about which part of it used to be a person. Some guesses are better. None are perfect.
Here is what’s on the table:
- Audacity’s Vocal Reduction and Isolation, the phase trick, now a separate plugin.
- Audacity’s OpenVINO plugins, which put an AI model and four stems inside Audacity.
- Demucs, Meta’s open-source separator for the command line.
- Ultimate Vocal Remover, a free desktop app with buttons.
- Spleeter, Deezer’s older and very fast separator.
Start with the best file you have. A low-bitrate MP3 ripped from an old video gives every model less to work with, and its artefacts get louder once the singer is gone.
Browser sites can remove vocals from a song too, but their free part usually ends at a preview. We ran six of them on one song in our stem splitter test, so this post sticks to tools you download and keep.
Audacity’s Vocal Reduction removes the middle of the song, not the singer
Audacity’s classic effect removes audio that is identical in the left and right channels. That’s all it does. The Audacity manual explains the basic method plainly: subtract one channel from the other, which is fast and returns mono. Lead vocals usually sit dead centre, so they cancel out. So do the bass, the kick drum and any solo instrument the engineer panned to the middle. In other words, the phase trick can remove vocals from a song only when the voice is the one thing in the centre.
The manual says as much. The effect “cannot distinguish” vocals from other centred sounds. Stereo reverb spreads across both channels, so it survives, and you get a ghost of the singer in an empty room. The input also has to be true stereo, not one mono signal copied twice. An Analyze action reports the channel correlation and pan position before you touch anything. Run it first. It tells you in two seconds whether the next ten minutes are worth it.
Simple mixes are the friendly case. Give it a voice in the middle, a guitar on one side, a piano on the other and not much reverb, and the phase trick sounds almost clever. Modern pop is not that.
Audacity’s sliders, its limits and the 3.5.0 change
Audacity’s effect has Low Cut and High Cut sliders that protect parts of the mix, and the default Low Cut is 120 Hz. That default suits low male voices. The manual suggests about 170 Hz for female voices and about 230 Hz for children. Anything below the cut stays in the mix, so the bass and kick survive. Raising the High Cut keeps cymbals and hi-hats. Sibilant “S” sounds can reach 5,000 to 8,000 Hz, though, and will then hiss on without their singer.
We tried the full-spectrum Remove Center option on a dense pop mix once. It did remove vocals from a song, along with the kick, the bass and the snare, like a band walking out over a pay dispute. That’s not a bug. Remove Center ignores the frequency sliders by design, so it takes everything in the middle.
Sliders help. But they don’t perform miracles.
There are practical limits as well. The effect works in memory, so selections longer than about half an hour can crash Audacity. Apart from the classic mono mode, it’s slow. And since version 3.5.0 it no longer ships with Audacity at all. It’s now a downloadable Nyquist plugin, and the same manual page says the newer OpenVINO AI plugins give better results. When the official manual recommends the competition, take the hint.
Audacity’s OpenVINO plugin can remove vocals from a song inside the editor
The OpenVINO AI plugins for Audacity add a Music Separation effect with four stems: drums, bass, vocals and other instruments. Intel builds them. The plugin’s own README says the separation uses Meta’s Demucs v4 model, ported to Intel’s OpenVINO toolkit.
The advantage is that it all happens where you already edit. According to the README, the features “run 100% locally” with no internet connection, on your CPU, GPU or NPU (a neural processing chip in newer laptops). You get four stems as tracks. Mute the vocal one, export the rest, done. To remove vocals from a song this way you need no terminal, no Python and no new app.
Two catches. Audacity’s download page lists installers for Windows and macOS only, while Linux users get build instructions and an afternoon. The same page also warns that the plugins aren’t yet compatible with the latest Audacity version. So check which release the plugin supports before you update anything. Otherwise you’ll install the editor twice.
The pack also includes noise suppression, Whisper transcription and a MusicGen generator. Those are a different post.
Separation on a laptop processor takes a while, so try it on a short selection before you commit a whole album to it and lose a second afternoon on top of the one the Linux build already took.
Demucs removes vocals from a song in one command, if you like terminals
Demucs is the open-source model under a lot of the tools in this post. Running it yourself takes one install line and one command. You need Python 3.8 or newer. The README gives the install as python3 -m pip install -U demucs. Then demucs --two-stems=vocals yourfile.mp3 writes a vocal file and an accompaniment file. That’s karaoke mode.
The default model, htdemucs, scores 9.00 dB SDR on the MUSDB HQ test set, per the README. SDR (signal-to-distortion ratio) measures how close a separated stem is to the real isolated recording. Higher is cleaner. A fine-tuned htdemucs_ft is slower and, in the README’s words, may be somewhat better. Then htdemucs_6s adds guitar and piano stems, although the README admits the piano one is weak.
Those scores come from songs that researchers have the real stems for, which is the only way to grade a separation. Your song is not in that set. Your song has a choir, a vocoder and a saxophone that sounds suspiciously like a person. Anyway.
Demucs needs around 3 GB of GPU memory at minimum and about 7 GB with default settings. That’s the hardware note people skip. The --segment option lowers memory use, and -d cpu runs it on the processor instead. Slower, but on any machine it will still remove vocals from a song. Two-stems mode still runs the full four-stem split underneath, so it saves no time or memory. It only saves you deleting three files.
The README also says Demucs is “not maintained anymore” because its author left Meta. Only important bug fixes go to his fork. The code is MIT-licensed and still works. Just don’t expect new features.
Ultimate Vocal Remover is the free app most people actually want
Ultimate Vocal Remover, usually called UVR, is a free, open-source desktop app at version 5.6. It puts several separation models behind a normal window with buttons. That makes it the easiest way to remove vocals from a song with a modern model and no terminal. The code is MIT-licensed. The README links a development update dated October 6, 2026, so the project is still moving.
UVR runs on Windows 10 or higher through an installer. A DirectML build covers AMD Radeon and Intel Arc cards. Mac users on Big Sur or later get separate builds for Apple Silicon and Intel, and Linux works through a manual install on Debian- and Arch-based systems. Install FFmpeg first. UVR needs it for anything that isn’t a WAV file.
Inside, you pick between three model families: MDX-Net, Demucs and the VR architecture. The core developers trained every bundled model except the Demucs v3 and v4 four-stem ones, the README says. So you can run one song through two or three models. Keep whichever result sounds least like it was recorded inside a fish tank. That comparison is UVR’s real advantage, because every song breaks a different model.
GPU conversion asks for an Nvidia GTX 1060 with 6 GB at minimum, and the README recommends 8 GB of video memory or more. AMD support is limited. The app is 64-bit only, and Pentium or Celeron chips aren’t guaranteed to work. Without a good graphics card it falls back to the processor, which can still remove vocals from a song, just slowly enough that you start it and go make dinner.


Spleeter is fast, older, and still fine for rough karaoke
Spleeter is Deezer’s open-source separator, and its selling point is speed: the README says it splits 4 stems “100x faster than real-time” on a GPU. It comes in three setups. Two stems gives vocals and accompaniment, four stems gives vocals, drums, bass and other, and five stems adds piano. It installs with pip and runs from the terminal, so it’s the same deal as Demucs.
Speed helps when you need to remove vocals from a song, then twenty more, for a whole setlist. It also helps for testing whether a song deserves a slower, cleaner pass in UVR. Spleeter is the older generation, though. Newer models in UVR came after it, and the gap shows on busy mixes. There, more of the singer stays smeared across the backing.
For a rough pass, that trade is fine. You hear whether the backing holds up without the voice, whether the bridge is just the singer and a pad, and whether the song is worth the slower tool at all. Fast first. Clean later.
Its README does something the others don’t. It tells you to get proper authorization from rights holders before running Spleeter on copyrighted material. Hold that thought.
Which free vocal remover fits which job
Pick the tool by where the instrumental goes next, not by the highest score. For singing along at home, any of them works. Audacity’s phase trick is fine on an old stereo mix with the voice dead centre. For a practice track, use UVR or Demucs and keep the vocal stem quiet as a guide instead of deleting it, the method in our karaoke version guide. To remove vocals from a song for a video, read the next section first.
Musicians get more from four stems than a missing singer. You can drop the guitar or the bass and play along with the rest; our post on making a guitar backing track from any song walks through it. For a remix, check the low end first. A kick drum often bleeds into the bass stem, as our stem separation guide shows.
Every AI tool that can remove vocals from a song leaves something behind. Expect a whisper of reverb where the voice was, cymbals that sound slightly underwater, and backing vocals that one model calls vocals and another calls “other”. If one model gives you warble, run the file through a second one before you blame the song.
We once ran one chorus through every model in the UVR dropdown, one after another, at two in the morning, and the honest result was two good stems, one fine one and several with strong opinions about the saxophone. Pick two. Compare. Sleep.
Removing vocals from a song does not remove the copyright
When you remove vocals from a song, the instrumental is still the original recording with a part taken out, so it belongs to the same people. The drums, bass line, chords and production all came from the master. Deleting the singer doesn’t make the rest yours. Taking the tyres off a car doesn’t make it a new car either.
Platforms know this. YouTube’s help page on Content ID says the system scans uploads against “a database of audio and visual files submitted by copyright owners”. On a match, the owner can block the video, monetize it with ads or track its viewers, and the result can differ by country. An instrumental cut from a song in that database is still mostly the same audio.
Singing your own vocal over it doesn’t change that. Your voice is new. The band isn’t.
Claims are common in general. Mubert’s own data, last 28 days: about one in four tracks run through our copyright checker came back with a rights holder’s claim (24%, down slightly from 25.9%). So check before you upload, whatever made the track.
If the singer was the only problem with your background music, an instrumental of the original won’t fix it. Mubert Render composes a new track from samples that real artists uploaded and get paid for, written to the exact length of your video. The licence scope depends on the plan, so read it like any licence. Our guide to what “no copyright music” actually means covers other places such music comes from.
A checklist to remove vocals from a song this week
- Run Audacity’s Analyze action to see if the vocal sits in the centre.
- Install FFmpeg before UVR if your files are MP3s.
- Check your graphics card memory against the 3 GB Demucs and 6 GB UVR minimums.
- Try
demucs --two-stems=vocalson one song, with-d cpuif there’s no GPU. - Run a difficult song through two UVR models and keep the cleaner one.
- Confirm the OpenVINO plugin supports your Audacity version before updating.
- Keep the vocal stem as a quiet guide track for practice.
- Check any instrumental of a commercial song for claims before it goes in a video.
The singer leaves with any of these tools. With an AI model, at least the bass player stays.



