What is stem splitting?
A stem is one isolated layer of a song: just the vocals, just the drums, just the bass. Stem splitting (also called source separation) uses a neural network to take a finished mix and reconstruct its layers — something that was mathematically impossible two decades ago, because a mixed track literally sums all instruments into one waveform.
The Demucs model powering this tool is trained to hear the difference between a voice, a kick drum, a bass guitar, and a piano, and to rebuild each as its own audio file.
How to split a song into stems
Upload the track and the AI returns up to seven stems: vocals, drums, bass, guitar, piano, other instruments, and the full instrumental. Preview each one right in the page and download only what you need. Free tier: 3 songs per day.
Demucs vs other stem separation tools
Demucs (developed at Facebook AI Research) is the open research model that most stem tools are built on — the quality you get here is the current state of the art for free tools. Proprietary services like LALAL.AI or Moises add tiered pricing and cloud accounts on top of similar (or the same) models. Professional DJ software like Serato Stems and VirtualDJ runs Demucs-class separation locally for live use.
This tool sends your file to the model once, returns the stems, and deletes your upload. No account, no subscription, no per-stem paywall.
Looking for UVR online?
UVR (Ultimate Vocal Remover) is the most respected free desktop app for stem separation — an open-source project that runs Demucs and other models locally on your machine. It's genuinely excellent, which is exactly why so many people search for "UVR online": they want the same results without the download, the setup, and a GPU.
This splitter is that online alternative. It runs the same class of Demucs models, works in any browser on any OS — Windows, macOS, Linux, even a phone — and needs zero installation or configuration. No model pickers, no settings files, no Python environment: upload the song, get seven stems.
The trade-off is control. UVR's desktop app lets you choose specific model checkpoints and tune inference settings, which matters when you're batch-processing thousands of files offline. For splitting a song or a dozen, the browser route gets you equivalent quality in less time than the installer takes to run.