Split a song into tracks in under a minute
When a song is mixed and mastered, every instrument is folded into a single stereo file. A song splitter reverses that, at least approximately. You give it the finished track and it hands back separate tracks for the main parts of the arrangement. This song splitter uses HT-Demucs, a hybrid transformer model from Meta AI Research, running directly on your graphics card through WebGPU. On a modern computer a typical song is split in 20 to 60 seconds.
You get four tracks that always line up with each other and with the original. Play them together and you hear the song; mute one and that part disappears. This makes the song splitter useful well beyond karaoke: for practice, transcription, remixing, mixing study and editing.
How to use the song splitter
- Drop the song file into the box above. MP3, WAV, FLAC, M4A, OGG and video files are accepted.
- On the first run the song splitter downloads its AI model (about 170 MB). It is cached, so later songs start immediately.
- Follow the progress bar. When it reaches 100%, the four tracks appear with their waveforms.
- Use play, mute, solo and volume to explore the song. Then download single tracks, all tracks as a ZIP, or your custom mix.
The tracks the song splitter produces
| Track | Instruments |
|---|---|
| Vocals | Lead vocals, backing vocals, choirs, rap, vocal chops |
| Drums | Acoustic kits, drum machines, cymbals, hand percussion |
| Bass | Electric and upright bass, synth bass, 808 and sub bass |
| Other | Guitars, piano, organ, synth leads and pads, strings, brass, sound effects |
What you can do with a split song
Learn your part
Solo the bass to hear every note of a bass line, or the other track to work out guitar chords and piano voicings. Then mute your instrument and play along with the rest of the band.
Remix and rearrange
Import the four tracks into a DAW and build something new: a slower piano version over the original vocal, a dance edit with new drums, or a stripped-down acoustic take. The drum remover and vocal remover pages offer one-click versions of the most common edits.
Study the mix
Hearing each part alone reveals how a professional mix is built: how much reverb sits on the vocal, how the kick and bass share the low end, how wide the guitars are panned.
Fix and edit recordings
Splitting a band rehearsal or a live phone recording lets you raise a buried vocal or tame a loud drummer after the fact. Results depend on the recording, but even rough splits are often enough for a usable balance.
How the song splitter works
The model looks at the song in two ways at once: as a waveform over time and as a spectrogram, a picture of which frequencies are playing at each moment. A transformer layer connects the two views so the model can use rhythm, pitch and timbre together to decide what belongs to which instrument. The song is processed in overlapping eight-second windows that are blended back together, so there are no audible seams.
Everything happens on your machine. The song splitter decodes the audio with your browser's built-in decoders, runs the neural network with ONNX Runtime on WebGPU, and builds the output files locally. Nothing is sent to a server.
Getting the best split
- Use lossless or high-bitrate files when you can; they give the song splitter more detail to work with.
- Modern studio recordings split more cleanly than old mono records or live recordings with a lot of room sound.
- Busy arrangements with many layered instruments will put more into the other track. That is expected.
- Keep the tab visible while it runs and close heavy apps; GPU memory is shared with everything else on screen.
Song splitter or stem splitter?
They are two names for the same job. "Stems" is the studio term for the separate tracks of a song, which is why our main page is called the stem splitter. This page uses the same engine and the same four tracks; it simply explains the tool for people who think of it as splitting a song into tracks.