This stem splitter separates a full mix into its parts — vocals, drums, bass and everything else — with a studio-grade neural network running entirely in your browser. Your music never leaves your device: the AI model is downloaded to you (once, then kept for future visits), not the other way around. The Studio quality model runs on your graphics chip through WebGPU and takes roughly one to three times the length of the song; the lighter Fast model does a vocals/instrumental split in a fraction of that time.
Common Use Cases
Remix and sample
Pull the drum break or the bassline out of a track and build on it. Stems export as WAV for lossless further production.
Practice with a band minus you
Mute the bass to play along yourself, or solo the drums to study a groove. Adjust each stem’s level to taste before exporting.
Make DJ tools and edits
Create drumless intros, acapella-out sections, or instrument-only versions for live blending.
Fix a mix you only have as a bounce
When the session files are gone, splitting the master lets you rebalance vocals against the band, then re-merge.
How to Split a Song into Stems Online
- Upload your song. The separation model starts downloading in the background as soon as you open this page.
- Open Split Stems in the AI section. Pick a model: Fast gives vocals + instrumental in a minute or two; Studio quality gives 4 stems — vocals, drums, bass and other — with the cleanest result.
- Click Split Stems and wait — with Studio quality this takes roughly one to three times the length of the song, with a live progress bar and time estimate.
- Mix the stems with the solo, mute and volume controls, then download them individually (WAV or MP3) or all together as a ZIP.
Common Problems and Fixes
The Studio model option says "Needs WebGPU"
The 4-stem model needs your browser’s WebGPU support, which every current version of Chrome, Edge, Firefox and Safari has. Update your browser, or use the Fast 2-stem model which runs anywhere.
Studio separation takes a long time
Running a studio-grade neural network locally is heavy — expect roughly one to three times the length of the song depending on your graphics hardware. The progress bar shows a live estimate; keep the tab in the foreground.
Cymbals or reverb bleed between stems
Reverb tails and cymbal wash overlap every instrument’s frequencies, so a faint print of them can remain in other stems. This is a limit of all current separation models, including paid ones.
Guitar and synths end up in the "Other" stem
The 4-stem model groups everything that isn’t vocals, drums or bass into Other. That is how the model was trained — a 6-stem model with separate guitar and piano is on the roadmap.
Frequently Asked Questions
How many stems can I get?
Two (vocals + instrumental) with the Fast model, four (vocals, drums, bass, other) with the Studio model. Pick the model inside the tool — your choice is remembered.
Which model should I use?
Fast for karaoke/acapella needs and slower devices — it usually finishes in a minute or two. Studio when you need drums and bass separated or the cleanest possible result, on a browser with WebGPU.
Is my song uploaded to a server?
No. Both models run inside your browser. Once the model has downloaded, even losing your connection mid-session won’t stop a separation.
How big are the downloads?
Fast: about 40 MB. Studio: about 199 MB including the GPU runtime. Both are cached on your device, so this happens once.
What formats can I export?
Each stem downloads as 16-bit WAV or 192 kbps MP3, and "Download all" packs every stem into a single ZIP.
Can I edit a stem afterwards?
Yes — the Edit button loads any stem straight into the editor, where every tool (trim, EQ, pitch, compressor…) applies to it. Undo restores the original full mix.