RemixStem

RemixStem / Audio guides

How to split a song into stems

Take a mix apart by ear. Learn how to isolate a vocal, build a backing track or pick out the bass line for practice.

By RemixStem · Updated

Inside a four-stem mixIllustration
Vocals
Drums
Bass
Other

One recording, four instrument groups. Choose vocals for an acapella, or combine drums, bass and other for a backing track.

Backing-track test

Which splitter keeps the bass?

Five local models processed the same recording. The low-end column shows the change in level from 40 to 150 Hz against the original backing. Closer to zero means a closer level match.

Five-second stereo excerpt · local CPU runs
ToolFree allowanceTest exportLow-end changeSoftware price
Demucs HTDemucsNo local quotaStereo WAV−0.01 dB$0
TorchAudio Hybrid DemucsNo local quotaStereo WAV0.00 dB$0
Open-Unmix umxhqNo local quotaStereo WAV−0.01 dB$0
Spleeter 2-stem FP16No local quotaStereo WAV−0.02 dB$0
UVR MDX-Net 9482No local quotaStereo WAV−0.01 dB$0

All five use your own computer. Compare all seven methods for the full-band results, or read how we measured them.

Use cases

What do you want to make?

Choose the part you want to keep and learn what to listen for after the split.


Three steps

How a song becomes four stems

  1. 01

    Start with a track you have the rights to use

    Your own recording, a track you licensed, or a buy-out beat. Most tools refuse or remove anything you cannot show a right to, and takedowns are common on anything taken from a streaming service.

  2. 02

    The model separates frequencies and timing

    Separation models are trained on multi-track recordings, so they learn what a vocal, a kick and a bass line look like in a waveform. The output is one file per instrument group, not a cleaned-up copy of the mix.

  3. 03

    You get one file per instrument

    Sing over the instrumental, keep the acapella, sample the drums, or drop the bass to practise a part. Fine-grained models add guitar and piano on top of the standard four stems.


A stem splitter takes a finished mix — vocals, drums, bass, guitars and keys flattened into one stereo file — and rebuilds it as separate tracks you can use on their own. That is the entire idea: one song in, four files out. Everything else on this page is about what those four files are actually good for, and where the result will disappoint you.

Stems, channels and multitracks are three different things

A multitrack is what a studio records: every microphone and instrument on its own channel, before anything is mixed. A stem is what a mixer exports for later use, usually a group of instruments such as all the drums together. A channel is one source on its own.

AI separation gives you something close to stems, but it is reconstruction, not recovery. The model has never seen the original session — it has learned what a voice, a kick and a bass line look like inside a waveform, then estimates how much of each frequency belongs to each source and rebuilds the tracks from that estimate. The output sounds like the song; it is not the same data the engineer had on the desk.

What the two model families give you

Split Tracks returned Good for
Standard four-stem Vocals, drums, bass, other Singing over a track, karaoke, sampling drums or bass
Fine-grained six-stem The four above, with guitar and piano pulled out of 'other' Isolating a guitar part, studying an arrangement

Always start with four. The fourth track, usually labelled other, holds guitars, keys, strings and anything else that is not voice, drums or bass. A six-stem model splits that track again, which is useful when you need one specific instrument and adds one more stage where artefacts can appear.

Where separation fails

Every stem splitter fails in the same handful of places, and knowing them saves you from blaming the tool for a bad source file.

  • Bleed. The instrumental usually keeps a ghost of the vocal, especially on loud, dense choruses. Quiet passages and backing vocals bleed most.
  • Cymbals and hi-hats. High-frequency percussion smears first, because it overlaps the frequency range of a voice.
  • Reverb tails. A vocal recorded with heavy reverb carries that reverb into its own file, and the instrumental keeps a tail of it as well.
  • Live recordings. Room sound on every microphone means every source is present in every channel, and the model has to guess harder.
  • Distorted or heavily processed vocals. Saturation and compression destroy the cues the model learned from clean recordings.

A clean studio mix separates dramatically better than a phone recording of a live show. That difference is bigger than the difference between two good models.

What it costs to run

The local tools in our backing-track comparison all ran on a CPU. They charge no per-song fee when you run them on your own computer. Hosted services set their own upload limits and export options.

The copyright part most guides skip

Separating a song does not change who owns it. You can process a recording you made, a track you licensed, or material released under a permissive licence. Pulling audio off a streaming service and splitting it produces a copy of a recording you do not own, and no tool makes that legal. If you publish the result, the platform's content matching will usually find it before a listener does.

What this site is for

RemixStem publishes practical guides to stem separation: what each job needs, which settings change the result, and how to check whether an output is usable before you build anything on top of it. Read the test method to see how we measured each backing track.

Start with the job you actually have: an acapella, a backing track, a karaoke version, or just the bass line.

Questions

Stem splitting questions people actually ask

What is a stem splitter?

It is a tool that takes a finished mix and returns separate tracks — normally vocals, drums, bass and a fourth group holding everything else. The tracks are rebuilt by a model, not recovered from the original session, so they are close to the source but never identical to it.

Is a stem splitter the same as a vocal remover?

Mostly yes. A vocal remover usually means the two-track job: voice out, instrumental left behind. A stem splitter names the same engine but asks for every group. If you only need a backing track, the vocal remover route is faster; if you also want the drums or the bass on their own, ask for the full split.

How many stems should I ask for?

Four covers almost every singing, karaoke and sampling job: vocals, drums, bass, other. Six-stem models add guitar and piano by splitting the 'other' track again, which helps if you need a specific guitar part, and adds a second chance for artefacts to creep in.

Can I separate a song from Spotify or YouTube?

Technically the file can be processed, but you would be making a copy of a recording you do not own, and that is a copyright problem in most countries regardless of the tool used. Work from your own recordings, tracks you licensed, or royalty-free material.

How good is AI stem separation, really?

Good enough to sing over and to sample, rarely good enough to pass as an original multitrack. Expect some vocal bleed in the instrumental, smeared cymbals, and reverb tails that follow the vocal into its own file. Live and heavily distorted recordings separate worst.

Do I need a DAW to use the files?

No for listening and singing along — any player opens them. Once you want to line tracks up, change levels or record a new vocal on top, a free DAW like Audacity or GarageBand is enough.

Where can I split my audio?

Choose an external separator from our comparison, then follow the guide for the stem you need. RemixStem publishes instructions and measured results; audio processing takes place in the tool you choose.