Stem splitter

What a stem splitter does

A finished song is one stereo file: every instrument has been mixed together. A stem splitter uses an AI model to pull it apart again, producing a separate track for each kind of sound. This one uses HTDemucs, the six-source version of the open-source Demucs model, which outputs vocals, drums, bass, guitar, piano and a sixth track for everything else (synths, strings, effects and so on).

Most online splitters upload your song to a server and charge credits for each one. Here the model runs in your browser with WebAssembly, on your own machine, so it’s free, private and unlimited. The trade-off is that your device does the work: the first run downloads the model (about 140 MB, kept for next time), and a split takes roughly as long as the song on a recent computer.

Every separation tool in one place

Times are for a typical four-core laptop. The extra tools download their own models (roughly 25–55 MB each) the first time you use them.

Mixing the stems

After the split, every stem appears in a small mixer with mute, solo and a volume control. Quick buttons cover the common mixes: Everything, No vocals (karaoke), Vocals only, Effects on vocals only and Effects on music only. If all you need is an instrumental or an acapella, the vocal remover page walks through that in more detail.

Because the stems always add up to the original, you can also use the splitter as a mix rebalancer: bring the vocal up a few dB on a muddy live recording, tame a loud kick, or turn down a busy accompaniment so the melody is easier to follow.

What people use stems for

Practising an instrument

Mute your own instrument and play along with the rest of the band, or solo it to hear exactly what the recorded player does. Combine it with the slow down settings to learn fast passages at a comfortable tempo without changing the key.

Remixes and edits

Each stem can skip or receive the effects. Put a big reverb on the vocals only, or make a slowed + reverb edit where the drums stay dry. For DAW work, solo each stem in turn and export it as 24-bit WAV or FLAC.

Studying arrangements

Hearing the bass line or the drum part on its own is the fastest way to understand how a song is built, whether you are transcribing it or teaching it. You can also turn any stem into a falling notes video to see the notes as well as hear them.

Cleaning up live recordings

Remove crowd & applause moves cheering and clapping into a separate track that you can mute or turn down, which is useful for bootleg concert recordings and school performances.

How stem separation works

Demucs was trained on multitrack recordings where every instrument was available on its own, so it learned what vocals, drums, bass, guitar and piano sound like, both alone and buried in a mix. HTDemucs, the version used here, looks at the song two ways at once: as a waveform (good for sharp drum hits and timing) and as a spectrogram (good for sustained notes and tone), and a transformer layer lets the two views share information. It then predicts each instrument’s part of the signal.

The song is processed in overlapping chunks of a few seconds, blended together so there are no clicks at the joins. That is why progress moves steadily rather than all at once, and why a long song needs more memory: all the separated tracks are kept in memory at full quality until you download them.

Tips for cleaner stems

• Use the best source. Lossless or high-bitrate files separate better. Low-bitrate MP3s have already thrown away detail the model relies on, which shows up as a watery sound on the stems.
• Pick only what you need. Untick stems before separating to keep memory down; anything unpicked stays in “Everything else”.
• Turn down rather than mute. When rebalancing a mix, small level changes sound more natural than removing an instrument completely, because any bleed stays in balance.
• Trim first for the slow tools. Lead/backing vocals and crowd removal take several times the song’s length, so run them on the part you actually need.

Limits worth knowing

• Memory: the main split needs about 1.3 GB of free memory. On phones and for songs over about ten minutes, untick the stems you don’t need, or trim the song first.
• Bleed: no separation is perfect. Piano and guitar can leak into each other, and reverb-heavy vocals leave faint traces in the instrumental.
• Beta tool: Pull out an instrument works best on clearly audible parts and leaks more than the main split.

Formats

Open MP3, WAV, FLAC, M4A/AAC, OGG, Opus, AIFF, WMA or the soundtrack of an MP4, MOV or MKV video. Export the current mix (or a soloed stem) as MP3 (128–320 kbps), WAV (16, 24 or 32-bit float), 24-bit FLAC, M4A or a 1080p MP4 video. Audio files have no watermark. The usual copyright rules apply: stems of a commercial recording are still that recording, so only publish remixes of music you have the rights to.