Split a recording into four stems, vocals, drums, bass and other, then play them back together, rebalance them, and export them as separate audio files. You can run more than one separation on the same recording and compare the results side by side.
What it does
Load a mono or stereo recording of up to three minutes, or open one of the saved demos, and press Split stems. StemSep shows the Original above and the Stemmed result below on one shared timeline, so every stem lines up with the source. Long recordings are processed in overlapping windows and stitched back together, and every stem keeps the full length of the recording.
Each stem has its own gain, balance, mute and solo, plus a download button. Several solos combine, and mute always wins. The mixer only affects what you hear: exported files are the raw separated stems, whatever the faders say.
Separation methods
The Model selector picks the method for the next run:
- Spleeter fp16 and Spleeter fp32 run on your own device, so your audio never leaves it. They are the same separation model at two levels of numeric precision, not two different algorithms. The model (about 79 MB or 157 MB) downloads the first time you use it and is then kept in your browser; Clear model cache removes it.
- Where the host offers them, SCNet, a second separation model, runs on a server, and Fadr is a paid service you connect with your own API key. Unavailable methods stay in the list with the reason they cannot run. Cancelling a Fadr job stops StemSep waiting for it, but the provider may still finish and charge for it.
Compare runs lists every finished separation with its method, where it ran and how long it took. Pick a row to listen to that result. The timings are a record of what happened, not a ranking of separation quality.
Exports and beat tracking
Download stems one by one as WAV files, or as a ZIP that also records which source, model and settings produced them.
Track beats in mix or Track beats in drums finds the beats and downbeats with a small beat-tracking model. Export beat grid MIDI writes them as a click on the drum channel (a side-stick, louder on downbeats) at their exact times. It is a grid of beats, not a transcription of the drum part, and it does not work out the song's tempo.
Limits
- Up to three minutes and 128 MB per recording.
- The four stems are mixed to one stereo output; there are no separate outputs per stem yet, so use the exports to work on stems individually.
- Long recordings and several retained results use a lot of memory.


