AI stem separation
An AI stem splitter that drops the stems straight into your mix
Separate any song into vocals, drums, bass and music on a cloud GPU in about a minute. Unlike standalone vocal removers, the four stems appear as lanes in your MixSongs project, ready to mute, solo, fade and export.
One click, four stems
What you get
-
Vocals
Lead and backing vocals, isolated with Mel-Band RoFormer, a state-of-the-art vocal model.
-
Drums
Kick, snare, hats and cymbals, from Meta’s fine-tuned Hybrid Transformer Demucs.
-
Bass
Bass guitar and sub-bass, separated from the instrumental.
-
Music
Everything else: guitars, keys, synths, strings and any vocal residue the first model left.
How the separation works
MixSongs runs two specialised models one after the other. First, Mel-Band RoFormer extracts the vocals. RoFormer models are among the best performers on public vocal-separation benchmarks and sound much cleaner than older “radio-like” separations. The instrumental that remains then goes to Demucs htdemucs_ft, Meta’s fine-tuned hybrid transformer, which splits it into drums, bass and other instruments.
The stems are written as lossless 24-bit FLAC files and are designed to add back up to the original song: anything a model was unsure about lands in the “music” stem instead of disappearing. So you can mute the vocals for a karaoke version, or solo them for an acapella, and the rest of the song stays intact.
Each separation runs on an NVIDIA L40S GPU and takes roughly a minute for a typical 3–4 minute song, including upload and download. A song can be up to 10 minutes long and the file up to 60 MB.
Separate a song in four steps
- 1
Add the song to MixSongs
Drag it onto the timeline like any other file. Mixing and editing stay free.
- 2
Click “AI” on the track
The separation dialog shows which stems you’ll get and how many credits it uses (one per song).
- 3
Sign in and get credits
Sign in with Google and buy a pack: 15 separations for $2.99, paid once through Paddle. No subscription.
- 4
Mix with the stems
The four stems appear under the track, each with a fader, mute, solo and its own volume checkpoints. Export the result or keep editing.
What people use it for
-
Karaoke versions
Mute the vocal stem and export an instrumental to sing over. See the karaoke guide.
-
Acapellas
Solo the vocals for remixes, mashups and practice. See the acapella guide.
-
Practice and transcription
Isolate the bass or drums to learn a part by ear, or turn them down to play along.
-
Better transitions
In a DJ-style mix, fade out one song’s drums before its vocals, or bring in the next song’s beat under the current vocal.
Your audio, handled carefully
Stem separation is the only MixSongs feature that sends audio over the internet. The file goes over HTTPS through our server to the GPU, and the stems come back to your browser. We do not keep your audio: it is deleted as soon as all stems have been downloaded, and an automatic sweep removes anything left within one hour. We never publish, share, listen to or train models on your files. Details are in the Privacy Policy.
Only separate audio that you own or have permission to use. Separating a commercial song does not give you any rights to it. Read the Terms and our Copyright policy before publishing anything you make.
Honest limits
| What to expect | |
|---|---|
| Quality | Excellent on most modern studio recordings. Heavily processed, live or very old mono recordings can leave artefacts or bleed between stems. |
| Stems | Four: vocals, drums, bass, music. Guitar and piano are not separate stems in the hosted app. |
| Length / size | Up to 10 minutes and 60 MB per song. |
| Speed | About a minute for a 3–4 minute song; the first job after a quiet period can take a little longer while the GPU starts. |
| Failures | If a separation fails, its credit is returned automatically. |
Stem separation FAQ
How much does it cost?
A pack of 15 separations costs $2.99 (taxes may apply depending on your country). It is a one-time purchase, not a subscription. One credit separates one song into four stems.
Is there a free trial?
The mixer, editor and export are free forever. Stem separation runs on paid GPUs, so it needs credits from the first song. A pack costs less than a coffee and failed separations are refunded automatically.
Is it a good vocal remover?
Yes. The vocal stem comes from Mel-Band RoFormer, one of the best publicly available vocal models. Muting it gives a clean instrumental for karaoke.
Can I download the stems separately?
In the app you can solo any stem and export it, or export your finished mix. Each stem behaves like a normal track.
Do you keep my songs?
No. Audio is deleted right after the stems reach your browser, and always within one hour. We never use it to train AI.