Stem Separation for Remixers: AI and Free Tools Compared

A finished song usually reaches you as a single stereo file. The vocal, drums, bass and every synth have already been mixed together, so you can't just turn the singer up or mute the kick. Stem separation (researchers call it music source separation) tries to undo that. Software analyses the mix and estimates what each part sounded like on its own, then writes those estimates out as separate audio files called stems.

For remixers this is a big deal. A clean acapella, an isolated drum groove or a bassline you can replace opens up remixes, mashups, DJ edits, practice tracks and sampling ideas that used to need the original multitrack session. This guide explains how AI stem separation works, compares the main free and built-in tools, and shows how to get the cleanest possible stems and use them well.

What stem separation actually does

When instruments are mixed, their sound waves add together. A vocal and a guitar that both sit around 1 to 3 kHz share the same frequencies at the same moment, so a simple EQ can't pull them apart. Older tricks, like phase-cancelling the centre of a stereo mix to remove a lead vocal, only work on some recordings, and they usually take the bass and kick drum with them.

Modern tools use neural networks trained on songs where the isolated parts are known. During training the model hears the full mix and learns to predict each source. At use time it applies what it learned to a song it has never heard. Ableton's manual describes the process as analysing "the spectral and temporal characteristics of an audio signal" and extracting the detected components into parts.

The result is an estimate, not a recovered original. Expect some bleed (traces of other instruments in a stem), artifacts (a watery or "underwater" texture, especially on reverb tails and cymbals) and occasional parts that land in the wrong stem. How much you hear depends on the song, the source file and the model.

Which stems can you get?

Most tools follow the layout of MUSDB18, the research dataset used to benchmark separation models. It contains 150 full-length tracks (about 10 hours) with isolated drums, bass, vocals and other stems, split into 100 training songs and 50 test songs. "Other" is everything that is not vocals, drums or bass: guitars, keys, synths, strings and effects.

That gives three common setups:

  • 2 stems: vocals and accompaniment (instrumental). This is "karaoke mode", the quickest way to get an acapella or an instrumental.
  • 4 stems: vocals, drums, bass and other. This is the standard for remixing.
  • 5 or 6 stems: some models also split out piano and/or guitar. They are useful, but these extra stems tend to be less clean than the core four.

How quality is measured (and why your ears still matter)

Research papers compare models with signal-to-distortion ratio (SDR), measured in decibels. Higher is better. It compares a separated stem with the real isolated recording. The Demucs project, for example, reports a score of 9.00 dB SDR on the MUSDB HQ test set for its Hybrid Transformer model. In the same comparison table, Spleeter scores 5.9 dB and Open-Unmix 5.3 dB. Demucs also publishes listening-test scores for quality and "contamination" (bleed), because a high SDR doesn't always mean a stem sounds natural.

For practical work, treat benchmark numbers as a rough ranking and then listen. Solo each stem, listen on headphones at a moderate level, and check the parts that usually fail first: sibilant "s" sounds, reverb tails, hi-hats and cymbals, and sections where the vocal and a lead instrument play the same notes.

Stem separation software compared

Here are the most widely used options. All of them can run on your own computer, so you don't have to upload a song to a website.

Demucs (free, open source)

Demucs is an open-source separation model originally released by Meta's research team. Version 4 uses a Hybrid Transformer design that processes both the waveform and the spectrogram. By default it separates drums, bass, vocals and other. A fine-tuned version (htdemucs_ft) trades speed for quality, and an experimental six-source model adds guitar and piano. The project notes that the guitar stem sounds "okay", but the piano stem has "a lot of bleeding and artifacts".

Demucs runs from the command line. demucs --two-stems=vocals song.mp3 gives a vocal and an instrumental, and --mp3 writes MP3 instead of WAV. If a GPU runs out of memory, the --segment option processes shorter chunks. You can also force CPU processing with -d cpu, which is slower but works on almost any machine. One caveat: the original repository says it is no longer maintained. Its author keeps a fork that only takes important bug fixes.

Best for: people comfortable with a terminal who want some of the best freely available quality.

Spleeter (free, open source)

Spleeter is Deezer's separation library, released with pretrained 2-stem, 4-stem and 5-stem (adds piano) models. It is written in Python on TensorFlow. Deezer says it can separate audio into four stems about 100 times faster than real time on a GPU. Spleeter's models were also built into commercial products, including iZotope RX 8's Music Rebalance, SpectraLayers 7's Unmix feature, VirtualDJ's stem isolation and Algoriddim's NeuralMix.

Newer models now beat Spleeter on separation quality. It's still a fast, lightweight choice for drafts or for processing many files in batch.

Best for: quick previews, batch jobs and developers building their own tools.

Ultimate Vocal Remover (free GUI)

Ultimate Vocal Remover (UVR) puts a point-and-click interface on top of several separation model families, including Demucs and MDX-Net models. Its installers bundle Python, PyTorch and the other dependencies, so you don't have to set anything up yourself. The core developers trained most of the bundled models themselves, apart from the Demucs v3 and v4 four-stem models. UVR offers Windows and macOS installers, and recent versions use GPU acceleration on Apple silicon Macs for Demucs v4 and MDX-Net.

Best for: free stem separation software without the command line, especially for vocal and instrumental extraction.

Ableton Live 12 (built in)

Live 12 has built-in stem separation that runs locally, with no internet connection or extra tools needed. Any mono or stereo audio can be split into Vocals, Drums, Bass and Others. You choose which ones to extract, and you can merge two or three of them into a single track. A High Speed mode is meant for drafts. High Quality mode runs dedicated passes for vocals, drums and bass and gets a higher SDR, but it takes longer. Ableton notes that older Windows PCs and Intel Macs use only the CPU, so separation can take a while. Cropping the clip to the section you need, such as just the chorus, saves time.

Best for: Live users who want stems to appear straight away as warpable clips in their set.

Logic Pro Stem Splitter (built in)

Logic Pro's Stem Splitter extracts vocals, drums, bass, guitar, piano and other instrumental parts. It can also split the drum stem into individual kit pieces: kick, snare, toms, hi-hat, cymbals and other percussion. Control-click an audio region and choose Processing > Stem Splitter. Then pick a preset or tick the stems you want. Any stems you leave unticked can be combined into a submix. Logic puts the results in a summing stack under the original track and mutes the original region. Stem Splitter needs a Mac with an M1 or later Apple silicon chip.

Best for: Logic users, and anyone who needs separate kick, snare and hi-hat stems for drum edits.

Online stem separation (free tiers)

Many websites offer stem separation online for free, often with limits on file length, number of songs or export quality. Many of them run open models like the ones above on their own servers. They are convenient on a phone or a low-powered laptop. Before uploading, read the terms on how long files are stored and whether uploads can be used for training. If you don't own a song, think twice before uploading it to a third party.

Which tool should you pick?

  • You already use Live 12 or Logic Pro: try the built-in option first. It is integrated, it runs locally and the stems are ready to edit.
  • You want the best free quality and don't mind setup: Demucs (fine-tuned model), either directly or through UVR.
  • You just need an acapella or instrumental quickly: a 2-stem mode in any of these tools.
  • You are processing lots of files: Spleeter or Demucs from the command line.
  • You are on a phone or an old computer: an online service, keeping the privacy caveat above in mind.

If you are still choosing a DAW, our guide to the best DAW for beginners covers which programs include tools like these.

How to get cleaner stems

  1. Start with the best source you can find. Use a lossless file (WAV, FLAC or ALAC) when you can. Low-bitrate MP3s and audio ripped from videos already have compression artifacts, and the model will carry them into every stem.
  2. Use the full song, not a re-recorded or speed-changed copy. Pitch-shifting, slowing down or adding effects before separation makes the job harder. Do those things to the stems afterwards.
  3. Choose the stem count you actually need. If you only need a vocal, a 2-stem split often sounds cleaner than taking the vocal from a 6-stem split.
  4. Use quality modes for the final version. Draft with the fast setting, then re-run the parts you keep in the higher-quality mode or model.
  5. Compare models on difficult songs. Dense rock mixes, heavily reverbed vocals and old mono recordings vary a lot between models. Running the same chorus through two tools and choosing the better result is normal practice.
  6. Check that the stems still add up. Play all the stems together at unity gain and compare them with the original. They should sound very close. If something is missing, it is often hiding in the "other" stem.

Cleaning up and using stems in a remix

Separated stems almost always need some extra processing. On an extracted vocal, a gentle high-pass filter removes leftover kick and bass rumble. A narrow EQ cut can tame a ringing piece of bleed, and a gate or careful clip editing removes noise between phrases. Low-level hi-hat bleed is often hidden by the new drums in your remix, so don't over-process before hearing the vocal in context. Our guide on how to mix vocals covers the EQ, de-essing and compression steps that also suit acapellas.

Some common ways to use stems:

  • Acapella remix: keep the vocal, then build new drums, bass and chords around it at a new tempo. The step-by-step process in our beginner remix tutorial works the same way with a separated vocal.
  • Drum and bass swap: mute the original drums and bass, keep the "other" stem for the song's harmony, and program a new groove.
  • Sampling: chop an isolated riff or drum break into a sampler. A clean stem makes a much cleaner sample than a chopped full mix.
  • DJ edits and mashups: combine the vocal of one song with the instrumental of another in the same key.
  • Practice and transcription: mute your own instrument to play along, or solo the bass to work out a line by ear.

Many online remix tools rely on the same stem separation. Our instant song remixer article explains why low-quality stems hold those tools back.

A note on rights

Stem separation is a technical step. It does not give you any rights to the music. Stems extracted from a commercial recording still contain that recording and the underlying composition. That's usually fine for practice, study or personal use. Releasing a remix, uploading it to streaming services or selling it normally needs permission from the rights holders. Official remix contests, label-supplied stems, stem-format releases and royalty-free sample packs are the safest material for a remix you plan to share.

Frequently asked questions

What is the best stem separation software? No single tool wins on every song. Demucs's Hybrid Transformer models are among the strongest freely available options on the standard benchmark. Ableton Live 12 and Logic Pro Stem Splitter are the most convenient if you already use those DAWs. Test two tools on a short section of your actual song.

Is there free stem separation software? Yes. Demucs, Spleeter and Ultimate Vocal Remover are all free to download and run on your own computer.

Can AI stem separation produce perfect stems? Not yet. Stems are estimates and usually contain some bleed or artifacts, especially in reverb, cymbals and parts that share a frequency range. In a busy remix, small flaws are often inaudible.

Do I need a powerful computer? No, but it helps. A GPU or an Apple silicon Mac is much faster. CPU-only processing works, but a full song can take several minutes. You can cut that time by processing only the section you need.

Can I separate stems online for free? Yes, many websites offer free tiers with limits. Check their file-storage and privacy terms before uploading.

Conclusion

Stem separation has turned finished songs into flexible material for remixers. Pick the tool that fits your setup, feed it the best source file you can find, choose only the stems you need, and judge the result by ear in context rather than by the numbers. Then clean up the stems lightly and get on with the creative part: the remix itself.

Sources

Tool features change between versions; check each tool's current documentation before you rely on a specific option.