Stem separation: how to pull vocals, drums and bass out of a song

What stem separation is

A stem is a part of the song that plays on its own: the vocals, the drums, the bass. In a studio they come from the session, grouping tracks before export. Stem separation is the reverse: starting from the finished mix, a single stereo file, and getting those parts back individually. It's what you're after when you want to remove the vocals from a song, pull out the acapella or separate the instruments to study them.

For years this couldn't be done well. The classic trick was to flip the phase of one channel and sum it with the other: whatever sat exactly in the centre disappeared. It removed the vocal, yes, but also the kick, the bass and the snare, which usually sit in the centre too, and it left any wide reverb untouched. EQ doesn't work either: as the sketch below shows, the instruments in a mix overlap across almost the whole spectrum, and a filter that cuts the vocal also cuts half the guitar.

If you want a refresher on which instrument lives in each zone, the frequency ranges chart goes through it band by band.

How AI does it

Today's separators are models trained on large numbers of songs whose tracks were available separately. They've learned what a voice, a drum kit or a bass sounds like, not which frequencies they occupy, and they use that to share out every moment of the mix among the stems. That's why they work where a filter fails: two sounds at the same frequency can go to different stems if they sound different.

One of the most widely used is Demucs, from Meta (Facebook Research). It's the one Fix The Mix uses, with its htdemucs model, and it splits into four stems:

The four stems come out the same length as the song and aligned with one another: summed at the same level, they give back something very close to the original mix. That's what lets you rebuild it with one stem missing.

What doesn't come out right

AI separation is good, not perfect, and it's worth knowing where it fails before you build anything on top. Soloed, almost every stem has something in it that shouldn't be there; inside a new mix, most of that residue disappears. The useful question isn't whether the stem is clean, but whether it's clean enough for what you're going to do with it.

• Bleed between stems. Leftover cymbals in the vocal, vocals in other, snare in the guitar. It's the most common issue and it's most noticeable soloed and in the gaps, when the main instrument stops.
• Watery or metallic sound. In very dense sections, with everything playing at once, the model hesitates and leaves artefacts: an odd sheen, watery or tinny, especially on the vocals and on other.
• The vocal's reverb and delays stay with the vocal. To the model they're voice, because they are. Don't expect a dry acapella from a song with a very wet vocal.
• Backing vocals and doubles go with the lead. If you only want the lead vocal, the vocal stem won't give it to you on its own.
• Shared low end gets split badly. A long 808, a low synth or a kick with a lot of tail can be split between bass, drums and other, sometimes switching stems in the middle of a note.
• The source mix calls the shots. A heavily compressed master, with everything glued together and no dynamics, separates worse than a mix with some air: the instruments reach the model more fused together.

The reverb issue makes sense: a reverb is the same voice repeated many times and attenuated, which is also why it can't be filtered out of a recorded take. The reverb and delay guide explains how that space is built, which makes it easier to decide what to put on top of an acapella that already brings its own.

What it's for

The most common uses, and what's worth knowing about each one:

• Instrumental or karaoke. Drums, bass and other together, without the vocal. If a vocal echo is left in other, you won't hear it live with someone singing on top.
• Acapella for a remix. The vocal stem, with its backing vocals and reverb. Put the new drums in before judging whether cymbal bleed is a problem: it's almost always covered.
• Practising your instrument. Remove the bass, the drums or the guitar (which will be in other) and play along with the original song. Practice doesn't need a perfect stem.
• Studying how a reference is mixed. Hearing the bass alone, the drums alone or the vocal alone from a record you like teaches you how much low end the bass carries, how much reverb the vocal has or how compressed the drums are, things you can't pick apart inside the mix.
• Remixing your own song when the tracks are lost. If all you have left of an old song is the bounce, separating it gives you something to work with: bring the vocal up, redo the drums or make an instrumental version.
• Sampling. Isolating a phrase, a drum loop or a bassline to use in another production, always with the rights to the original.

The fourth use is the one that teaches the most. Hearing a reference's vocal on its own and then going back to yours changes what you hear in both; the reference tracks guide covers how to compare without fooling yourself, and the one on kick and bass helps you make sense of what you hear when you isolate the bass and drums of a record that sounds good.

How to get the best stems

The model does what it can with what you give it, and what you give it matters more than it seems:

• Upload the highest quality file you have. A WAV or FLAC separates better than a 128 kbps MP3: lossy compression has already erased high-end detail and smeared everything a little, and that's exactly what the model needs to tell a cymbal from an S.
• If the song is yours, upload the mix, not the master. The master carries more compression and limiting, and the instruments arrive more glued together. The mix, with its dynamics, separates more cleanly.
• In your DAW, place the stems from 0:00. They come out the same length as the song and aligned, so if they all start at the same point they play in place without moving anything.
• Before processing a stem, sum it with the others and compare against the original song at the same level. If it sounds the same, the separation is fine and what you hear soloed is bleed the mix covers.

A separated acapella is then treated like any vocal take, with one difference: it already comes EQ'd, compressed and with effects. Better to take away than to add, and go carefully when boosting the top, which is where cymbal bleed lives. The vocal mixing guide picks up from here.

Try it on your song

Everything in this guide is what the online stem separation tool does: you upload a song up to 7 minutes long and it splits it with Demucs into vocals, drums, bass and other. In a real test, a 2:27 song took 32 seconds from start to finish. Each stem can be soloed or removed from the mix to hear the rest without it.

• Free, no account: you split 30 seconds of your song and hear each stem with solo and remove.
• With Pro: the whole song, stem downloads as 24-bit WAV and up to 20 separations a month.
• Privacy: unlike most of Fix The Mix's tools, which run in your browser, this one needs a server. The song is uploaded to our server and processed by Replicate, in the United States. The song and the stems are deleted automatically after 30 days.

And if you want to play with them, in Mix Space you can place your stems in space: drag each one left or right, bring it forward or push it back, and hear how the mix changes as you move them.