How to use Stem Splitter
- Choose a song (MP3, M4A, WAV, FLAC, OGG or a video) on your phone or computer, or try the 15-second example.
- Keep "4 stems" selected (or switch to vocals and instrumental if that is all you need).
- Tap "Split into stems". The AI downloads once (54 MB) and then works on your device, about eight seconds of song at a time, with the time left shown.
- Listen to each stem, then save the ones you need as WAV files or all four in one ZIP.
Why use Stem Splitter
- Four stems from one file
- Drums, bass, vocals and "other" (guitars, keys, strings, synths, backing pads): the standard four-stem split that remix packs and DAW sessions use.
- On your device, nothing uploaded
- The whole separation runs in the browser. Only the model (54 MB) downloads, once, from this site.
- Stems you can hear first
- Every stem has its own waveform and plays on its own, so you know what you are saving before you save it.
- Studio-grade model
- Hybrid Transformer Demucs (Meta, MIT licence), the open model behind many paid stem services, exported to run in the browser with the same maths.
- WAV, ready for any DAW
- 44.1 kHz 16-bit stereo files named after the song and the stem, so they drop straight into Ableton, FL Studio, Logic, Audacity or a karaoke app.
- Stop and keep
- Stop a long song early and the stems of the part that is done are kept.
About this tool
Producers, DJs, music students and cover bands all need the same thing from a stem splitter: the parts of a finished song, separated well enough to work with. A drummer wants to hear the kit alone to learn a fill; a bassist wants the bass line without the guitars over it; a DJ wants the vocal for a mashup and the drums for a loop; a teacher wants the melody without the singer. Until recently that meant a paid service and an upload. This page runs the same kind of model, Demucs from Meta research, inside the browser. The song is cut into pieces of about eight seconds; each piece is turned into a spectrogram, passed through the model, and turned back into sound, and the pieces are joined with a cross-fade so nothing clicks at the joins.
The split is "four stems" because that is how the model was trained: drums, bass, vocals and other. Anything that is not one of the first three goes into other, so a piano, a guitar solo, a harmonium or a string section all land there together. The stems add up to the original song, and each comes out as a stereo WAV at full quality. Expect a little bleed on busy passages, where a crash cymbal can leave a whisper in other or a very low synth can share with the bass; that is normal for any separation model, including the paid ones. The tool is for music you own or have the right to use, such as your own recordings, stems you need for a licensed cover, or royalty-free tracks. A phone can split a song of up to six minutes; a laptop takes longer songs and finishes in about twice the song's length.
Frequently asked questions
Is this stem splitter free?
Yes, with no sign-up or watermark. The AI runs on your device, so there is no per-song cost; the model downloads once (54 MB) and is kept in the browser.
Is the song uploaded anywhere?
No. The file is decoded and separated in your browser. We never receive it; only the model comes from this site.
Which stems do I get?
Drums, bass, vocals and other, each as a 44.1 kHz 16-bit stereo WAV named after the song. Together they add up to the original mix.
Can it separate a guitar or a piano on its own?
Not as a separate stem: the model was trained on four classes, so guitars, keys, strings, synths and traditional instruments all go into "other". If you need only the voice and the backing track, the vocal remover page gives those two files.
How long does it take, and does it work on a phone?
About twice the length of the song on a recent laptop; a phone takes a few minutes for each minute of music and accepts songs up to six minutes. The screen stays awake while it works, and the progress bar shows the time left.
Can I use the stems in Ableton, FL Studio or Logic?
Yes. They are plain WAV files at 44.1 kHz, the same length as the song, so they line up when dropped on parallel tracks at the same start point.
Does it work on Bollywood, Coke Studio or Arabic songs?
It runs on them the same way: the model listens for the sound of drums, bass and voices rather than a language or style. Tabla and dholak go to drums; harmonium, sitar, oud and strings go to other. We have measured the split on English pop so far; layered harmony voices can partly land in other.
What about copyright?
Use it on music you own or may use. Splitting a commercial record for your own practice is your call; publishing or selling the stems needs the rights holder's permission.
Built and maintained by Saad Naeem
AI engineer, Doha · Updated