Split Any Song Into Six Stems

What a stem splitter does

A stem is one layer of a finished song: the vocal, the drum kit, the bass. Producers have the stems of their own tracks, but the public almost never does. A stem splitter works backwards from the stereo file, estimating each layer with a neural network so you can pull a song apart after the fact.

This page runs the 6-stem Demucs model. Compared with the common 4-stem split it separates guitar and piano from the leftover instruments, which matters for rock, pop and singer-songwriter material where those two carry the harmony.

Where six stems beat four

With four stems, a track with a rhythm guitar and a piano gives you one crowded other stem. With six, you can mute the guitar and play it yourself, or isolate the piano to transcribe the chords. Bands use it to build custom practice tracks, and DJs use the drum and bass stems for transitions.

The trade-off is accuracy. Guitar and piano are harder for the model than vocals and drums, so expect a little bleed between them and the other stem. If you only care about a clean voice and rhythm section, the four-stem preset on the vocal remover page is a more conservative choice.

Working with the files afterwards

Download the stems as WAV if you will edit them, since each MP3 re-encode adds artefacts. Import everything into your DAW at the same start position and the stems will line up exactly, because all of them are rendered from the same timeline.

From there, common next steps are tempo and key detection on the isolated bass, converting a piano or guitar stem to MIDI, or sending the vocal stem to a mastering chain. Stems are raw material, so a gentle high-pass filter on the non-bass stems often cleans up low-frequency bleed.

Limits worth knowing

Six stems is the ceiling of this model: there are no separate stems for strings, synths or backing vocals, and everything the model cannot assign lands in other. Very busy mixes with several guitars or layered keyboards will blur between guitar, piano and other. Files are capped at 100 MB and 10 minutes, and the separation is deterministic enough that running the same song again gives a very similar result, so spend your credits on a better source file rather than on retries.