Stem separation is software that takes a finished, mixed song and pulls it apart into separate tracks, usually 4 of them: vocals, drums, bass and “other”, which means everything left over. If you want to remix a song but only have the MP3, this is how you get parts you can move around in a DAW (a digital audio workstation, the program you make music in: Ableton Live, Logic, FL Studio and the rest). This post covers the file to start from, how to split it, how to load the stems so they stay in sync, and how to clean up the parts the model got slightly wrong.
The slightly wrong parts come up a lot, so here is the main thing to expect. The stems sound very good, with the occasional kick drum hiding in the bass.
Stem separation gives you four guesses, not the original session
A split song gives you four stereo files, each the length of the whole song. They aren’t the producer’s original multitracks. A model listened to the mix and estimated what each instrument probably sounded like on its own. Usually the estimate is good. It’s still an estimate.
Four stems is the standard for stem separation almost everywhere. Meta’s open-source Demucs is the research model much of this field grew out of. It separates “drums, bass, and vocals from the rest of the accompaniment” into four files. Ableton’s Live 12 manual uses the same four. Its Others stem holds whatever wasn’t detected as vocals, drums or bass.
Some tools add guitar and piano. Demucs calls its own six-source model experimental and says plainly that the piano stem has “a lot of bleeding and artifacts”. Bleed is when one instrument leaks quietly into another instrument’s stem, and it’s the mess stem separation leaves you to clean up most.
Things that share a frequency range fight hardest. Our Stem Splitter page says it directly: vocals, drums and bass “usually separate cleanly”, but bass and kick drum “can bleed into each other a little.” That’s the raisin in the cake. You’ll learn where it hides.
Stem separation only hears what’s left in the mix
A model can’t recover sound that the file has already thrown away. A 128 kbps MP3 ripped from a video is missing high-frequency detail, so the model has less to work with. The stems come out smeared. Hi-hats suffer most, because they live up in the frequencies lossy compression cuts first.
We tried this once with a ripped file from an old phone. The vocal stem was usable. The hi-hats sounded like bacon frying at the far end of a tunnel. Then we split the lossless version of the same song, and the bacon turned back into cymbals.
So feed stem separation the best file. It has to be one you can legally get. A WAV or FLAC bought from the artist’s store beats a streaming rip, and a streaming rip beats a screen recording. Our splitter accepts MP3, WAV, FLAC, M4A and MP4 video, which is handy when the only copy you have is a clip. The output quality still follows the input quality.
Loud masters are a quieter problem. A heavily compressed mix has instruments squashed together in time as well as frequency, so transients (the sharp front edge of a drum hit) blur into the vocal. Separation still works on these songs. You’ll just do more cleanup.
Pick how many stems you need before you pick the tool
For most remixes four stems are plenty. More stems from the same stem separation pass mostly means more bleed, so work out what you’re keeping before you upload anything.
The common remix plans each need different parts:
- An acapella remix keeps the vocal stem and replaces the rest, so you only need one clean stem.
- A rework keeps the drums and bass and rebuilds the harmony, so you need the rhythm section clean.
- A flip keeps one hook from the “other” stem, which means that stem matters most.
- An instrumental drops the vocal, so the vocal stem only has to be good enough to delete.
The acapella case is easiest because voices are what models are trained on hardest. We covered the reverse trick, keeping the band and turning the singer down, in our post on making a karaoke version of any song. For guitar players, taking the guitar out and keeping the band uses the same single-instrument approach.
On the Mubert Stem Splitter you can pick “All stems (4)”, or pick one instrument (drums, bass, guitar or piano) and choose Remove or Extract. Either way you get that part plus the song without it. Six-stem splits with separate guitar and piano are on paid plans. A free account gets four-stem splits with full-length MP3 downloads and 20 minutes of songs a month. Without an account you hear a 30-second preview.
Run stem separation in the browser or inside your DAW
There are two ways to do it. A browser tool splits the song on someone else’s computer and gives you files. A DAW with built-in separation does the job on your own computer, inside the project. Both work. The difference is mostly which hardware and which edition you already own.
Here is what the big DAWs offer, from their makers’ own pages:
- Ableton added stem separation in Live 12.3 on November 25, 2025. It “runs locally” and comes with Live Suite only. You get four stems, plus a High Quality mode that runs a dedicated pass each for vocals, drums and bass and builds Others from what’s left.
- Apple’s Logic Pro Stem Splitter got guitar and piano stems in the May 28, 2025 update. It “requires iPad or Mac with M1 chip or later.”
- FL Studio added built-in stem separation in version 21.2 in 2023.
Ableton’s manual also warns that processing can take several minutes on older machines. Cropping the clip to the section you need first saves time. That’s sensible advice, and it brings up a sync problem we’ll get to next.
The browser route makes sense when your DAW doesn’t have separation, when your laptop is a 2017 Intel machine that sounds like a hair dryer under load, or when you want to compare a few tools on the same song. We ran that comparison on one track in our stem splitter test of six tools and their free previews. In our splitter you can also mute, solo or turn down any stem in the browser before downloading, so you can check the split before it reaches your session.
Line every stem up at the same start point
Stems only stay in sync if they all start at exactly the same sample. That’s the whole trick. It’s also the step most people get wrong on their first remix.
Every file a stem separation tool exports is the full song length and starts at the same moment as the original. Keep it that way. Select all the stems, drag them into the DAW in one go, and put them on separate tracks starting at the same point, ideally bar 1. If one stem lands half a beat late, the whole remix sounds like the drummer is in a different room. Technically he is.
Two more settings cause most sync accidents. First, match the project sample rate to the files. Demucs, for example, writes stereo WAV at 44.1 kHz. Drop that into a 48 kHz project and, depending on the DAW, it either resamples on import or quietly plays at the wrong speed. Second, switch off automatic time-stretching (Ableton calls it warping) until all the stems are in. If each file gets its own tempo guess, four files can end up with four slightly different guesses.
Once everything is in, find the original song’s tempo, set the project to it and line the grid up with the first downbeat. Only then turn stretching back on and drag the tempo to wherever your remix lives. All the stems move together because they all follow the same grid.
Cropping before you split, as Ableton suggests, is fine. Crop all the stems the same way, though, or crop the original before you split it. If you trim the vocal by hand and not the drums, you’ve started a sync puzzle that will cost you an evening. We spent one. It wasn’t fun.
Check the stem separation by putting the stems back together
Stem separation has a built-in quality check. Play all the stems together at 0 dB with no effects and they should sound almost exactly like the original song. If they don’t, something went wrong in the export or the import.
The stricter version is a null test. Put the original song on its own track and flip its polarity (the “phase invert” button on most channel strips). Play it with all the stems. Wherever the stems add up to the original, the sound cancels out, and what’s left is whatever the separation lost or added. In Ableton’s High Quality mode, Others is built from the audio remaining after the other passes, so the sum should come close. Tools that model every stem separately can leave more residue.
The sum test tells you whether the files are intact. To judge each stem, solo it and listen. Expect the vocal to carry a faint ghost of the snare, the bass to thump on every kick and the “other” stem to sound watery around the edges. The Demucs authors put a number on quality: their v4 model scores 9.0 dB SDR on the MUSDB HQ test set. SDR (signal-to-distortion ratio) measures how close a stem is to the real isolated part, and higher is better. In practice that means very good, but not invisible.
Anyway.
Clean the vocal and drum stems after stem separation
Bleed is easier to hide than to remove. Hiding it is usually enough. A remix adds new drums, new chords and new effects, and those cover a lot of small mistakes, so clean each stem only as far as the remix actually exposes it.
On the vocal stem, a high-pass filter below the singer’s lowest note takes out most of the kick and bass ghosts. A gentle de-esser can soften the swishy artefacts that separation leaves on “s” sounds. If you’re going to drench the vocal in reverb, which half of all remixes do, you can stop there, because the reverb hides the rest.
On the drum stem, a gate (an effect that mutes audio below a set level) between hits removes the faint vocal wash. Set it too hard and the cymbal tails get cut off. If you only need the kick and snare, a low-pass filter on a duplicate of the stem gives you a cleaner kick than a gate does.
Bass against kick is the hardest bleed to fix
Bass against kick is the raisin problem. A dynamic EQ or a sidechain compressor on the bass stem, triggered by the drum stem, dips the bass for a moment on each kick, so the leftover kick inside the bass gets ducked along with it.
The other option is to keep the bass stem only as a reference. Play the bassline again yourself on a synth and mute the original. Plenty of producers do this, and it’s often the cleaner result.
The “other” stem is the least cleanly separated by design, because it’s where the model puts everything it couldn’t name. Use it for texture, in short chopped sections or low in the mix. Built-in stem separation was a big deal for exactly this reason. The DAW used to be the last stop, and now the splitting happens inside it, which is also why stories like Suno’s Bridge beta putting AI stems inside Ableton Live get so much attention.
Stem separation is legal to run, but a released remix needs permission
Running stem separation on your own laptop to practise is one thing. Releasing the remix is another. US law gives the copyright owner the exclusive right to “prepare derivative works based upon the copyrighted work” (17 U.S.C. § 106), and the legal definition of a derivative work explicitly includes a musical arrangement and a sound recording that recasts the original (17 U.S.C. § 101). A remix that uses someone’s actual recording involves two owners. One holds the composition (usually a publisher), the other holds the recording (usually a label).
Neither the split nor the cleanup gets you a licence. Your EQ work is your own, but the vocal under it still belongs to someone else, and content ID systems recognise vocals very well. In Mubert’s own data from the last 28 days, about one in four tracks run through our copyright checker came back with a rights holder’s claim (24.1%, down slightly from 26.8%). That’s a lot of people finding out after the upload rather than before.
The safe routes are dull but real. You can remix tracks whose owners released stems for exactly that purpose, enter official remix contests, ask the artist directly (smaller artists say yes more often than you’d think), or keep the remix private as practice. For a background or bed under your own vocal, you can generate an original track at the exact length you need, which avoids the two-owners problem entirely. Our explainer on what “no copyright music” really means covers where cleared music actually comes from.
This is a summary, not legal advice. Licence scope depends on the deal and the country. If the remix is going to a label or a distributor, have someone with a law degree read the paperwork.
A checklist for your first stem separation remix
- Buy or find the best available file of the song, WAV or FLAC if possible.
- Decide which stems the remix keeps before you split anything.
- Split it into four stems first, and only go to six if you really need guitar or piano on their own.
- Set the project sample rate to match the stem files before importing them.
- Drag every stem in together, starting at bar 1, with time-stretching switched off.
- Check the split by summing the stems against the polarity-flipped original.
- Clean only the bleed the remix exposes, starting with the bass against the kick.
- Get permission from both owners, or keep the remix private, before uploading it anywhere.
Then solo the bass stem one last time. If you can still hear the kick, you know where to put the sidechain.
Who we are
