Drop a Song Here
MP3, WAV, FLAC, OGG, M4A — video files work too
Four Stems From One Song
Drop a song in and it comes back as four tracks: vocals, drums, bass, and everything else — guitars, keys, synths, strings. Four neural networks read the song's spectrogram and each one estimates how much of every frequency, at every moment, belongs to its instrument. The four tracks are the same length as the original and stay in sync with it, so they load into a DAW as a session that plays back as the song.
What the Fourth Stem Is
Three of the four are self-explanatory. The fourth, labelled Other, is everything the model was not trained to name: guitars, piano, synths, horns, strings, percussion that is not the kit. It is not a leftovers bin in the sense of being noise — on most records it is the harmonic body of the arrangement, and on a guitar band it is the loudest stem of the four.
What Stems Are For
A remix starts with the parts. A drum stem gives you the groove to build a new track under, a bass stem gives you the line to replace or re-amp, and the vocal stem is the one that has to stay. Practising musicians use the inverse: mute the bass stem and play along with a real band. And in video work a music bed often needs the drums pulled down under dialogue, which is a two-second job with stems and impossible without them.
It Takes Longer, and Here Is Exactly How Long
Four stems means four models instead of two: 157 MB to download the first time against 78, and about three and a half minutes of work for a four-minute song against roughly one minute. Both numbers are measured on a normal laptop, on one processor thread, with no graphics card involved. The model download happens once and is then cached on your device; the separation time is per song. If a two-way split is all you need, the vocal remover is the faster tool by a wide margin.
Open the Vocal Remover →The Song Never Leaves Your Device
Every other free stem splitter uploads your song and separates it on a server, which is why they meter you by track length, by tracks per day, or behind an account. Here the model comes to the song instead: it is downloaded to your browser and the audio is never sent anywhere. You can verify that in the browser's Network tab — after the model arrives, nothing outbound happens at all.
Mix the Stems Back Together
All four tracks land in the audio pool, so the obvious next step needs no downloading: send them to the Mixer, drop the stem you want gone, and export the rest as one file. That is how a backing track without the bass, or a drums-and-vocals-only version, gets made in about a minute.
Open the Audio Mixer →Vocals · Drums · Bass · Other
Four tracks from one song, full length, in sync. Download any of them, or all four.Nothing Is Uploaded
The song stays on your device — verifiable in the browser's Network tab. No account, no track limit.157 MB, Once
The four models download on first use and stay cached. About three and a half minutes per four-minute song after that.Frequently Asked Questions
Four: vocals, drums, bass, and other. Other is everything that is not one of the first three — guitars, keys, synths, strings, extra percussion. All four are the same length as the source and stay in sync with it, so they open in a DAW as a working session.
About three and a half minutes for a four-minute song on a normal laptop, plus a one-time 157 MB download of the four models. It runs on one processor thread and needs no graphics card. A two-way vocal split is roughly three times faster if that is all you need.
No. The models are downloaded to your browser and the separation happens on your device. Open the browser's Network tab to confirm it: the models arrive on first use, and after that there is no outbound request at all.
Good enough to remix with and to practise against; not a substitute for the original multitrack. Drums and bass come out cleanest because they own their frequency ranges. The Other stem collects what the other three could not claim, so artefacts concentrate there.
Not here. The model this page uses is trained on exactly these four categories; separating a guitar from a piano is a different model and a different tool.
Ten minutes per file. Lower than the hour the other tools allow, because separation holds the whole song in the tab as uncompressed audio plus one buffer per stem — with four stems that is roughly 190 MB per minute of audio.
Spleeter's four-stem model, published by Deezer under the MIT licence, exported to ONNX and run with onnxruntime in your browser. The licence is why it was chosen over the research models that score better: those are released for scientific use only or under non-commercial terms.