A free stem splitter that runs in your browser
Most online stem splitters upload your song to a server. Stemina brings the model to your browser instead, so the audio never leaves your computer.
What it does
Drop in a song and Stemina returns six separate stems:
- Vocals — lead and backing vocals.
- Drums — kit and most percussion.
- Bass — bass guitar and synth bass.
- Guitar — acoustic and electric.
- Piano — acoustic and electric piano.
- Other — everything else: synths, strings, pads, effects.
Then it gives you a small studio to work with them: all six lanes on one clock with solo, mute and gain; a Slice bench for cutting samples; BPM detection; and exports as 16-bit or 32-bit float WAV, a ZIP of every stem, or an Ableton Live project with each stem already on its own track.
How it is different
- Nothing is uploaded. There is no server to send audio to. The separation runs in a background thread of your browser tab.
- Free, without quotas. No account, no credits and no daily cap. A single track can be up to 30 minutes long. Stems are written to disk as they are made, but the decoded source stays in memory, so memory grows with length: about 64 MB of source audio for a three-minute song and about 635 MB for thirty minutes.
- Batch processing. Add several tracks; they run one at a time from an editable queue, and a failed file can be retried without stopping the rest.
- Works offline after the first run. The model is cached on your device, and finished separations reopen in about a second after a reload.
- No accounts, no tracking. The app ships no analytics of any kind.
What it needs, honestly
| You have | What to expect |
|---|---|
| Desktop Chrome or Edge 121+ with WebGPU | The fast path. The model runs on your graphics card. Speed depends on the GPU: an integrated Intel Iris Xe takes about 17.6 s per four-second segment; a dedicated card is much faster. |
| A browser without WebGPU | Still works, on the processor through WebAssembly, roughly ten times more slowly. Our 15-second sample took just under two minutes. |
| A phone or tablet | May start, but is unlikely to finish a full song: phones throttle under sustained load, and the work stops when the screen locks. |
The first run downloads the model: 353 MB on most machines, or the 701 MB full-precision version on a GPU without half-precision support. Keep about 1 GB of storage free.
The model
Stemina runs BS-RoFormer-SW, a band-split transformer trained for six-stem separation by the open-source music-separation community, exported to ONNX and executed with ONNX Runtime Web. The song is resampled to 44.1 kHz stereo, cut into four-second segments, turned into a spectrogram, separated, turned back into audio and cross-faded so the joins are inaudible. Nothing post-processes the result: no de-noiser, no de-clicker, no filtering.
Separation is the model's judgement, not the original multitrack. Expect some bleed between stems, especially reverb, doubled parts and loud drum transients.
When an online service may suit you better
If you only have a phone, or a very old computer without WebGPU and a long track, a server-based splitter will usually finish faster, because the work happens on someone else's hardware. The trade is that your audio goes with it.
Last updated