How to separate vocals from a song

The short version

1. Open stemina.bpmtech.no in Chrome or Edge on a desktop or laptop.
2. Drop in the song. MP3, WAV, FLAC, M4A and OGG all work. You can add several files and they will run one after another.
3. Let the model download once. About 353 MB on most machines. It is stored in your browser, so later runs start straight away and work offline.
4. Press Separate. The track is processed in four-second segments on your graphics card, or on the processor if the browser has no WebGPU.
5. Solo the vocals in the Mix bench to check the result.
6. Export the vocal stem as a 16-bit or 32-bit float WAV, all stems as a ZIP, or an Ableton Live project with every stem on its own track.

What you get

One pass produces six stems at 44.1 kHz stereo: vocals, drums, bass, guitar, piano and other (synths, strings, pads and anything that is not one of the first five). The six line up sample for sample, so they drop into a DAW at bar 1 and play back as the original mix.

Want the instrumental instead of the vocal? Mute the vocal lane in the Mix bench to hear it. Muting changes playback only: the ZIP and the Ableton project always contain every stem, vocals included. So export all stems, then delete or mute the vocal track in your editor and combine the other five. Stemina exports stems, not a single bounced instrumental file.

Before you start

• Browser. Chrome or Edge 121 or later with WebGPU is the fast path. Without WebGPU the app falls back to WebAssembly on the processor and still finishes, roughly ten times more slowly.
• Time. It depends almost entirely on the graphics chip. The slowest GPU we have profiled, an integrated Intel Iris Xe, takes about 17.6 seconds per four-second segment. At the default 25 % overlap each segment advances three seconds, so a four-minute song is 80 segments: roughly 23½ minutes of separation, plus loading the model. On the WebAssembly fallback our 15-second sample took just under two minutes.
• Storage. Keep about 1 GB free for the cached model and the working files.
• Phones. A recent phone may start a separation, but heat, memory and the screen locking make finishing a whole song unlikely. Use a computer.

Getting the cleanest vocal

• Start from the best copy you have. A WAV or FLAC, or a high-bitrate MP3, gives the model more to work with than a low-bitrate download.
• Raise the overlap for less hiss. The Configure screen lets neighbouring segments share up to 50 %. Going up buys roughly a decibel less broadband hiss and costs proportionally more time.
• Expect some bleed. Reverb tails, doubled backing vocals and vocal-like synths can land in either stem. A short tick in the vocal is usually a drum hit the model filed in the wrong place; no filter removes it cleanly, so cut around it in the Slice bench.
• Nothing is cleaned up after the model. Stemina applies no de-clicker, no noise reduction and no normalisation, so what you hear is exactly what the model produced.

Cutting a vocal sample

Switch to the Slice bench, pick the vocal stem, zoom in and drag across the phrase you want. Cuts snap to zero crossings and get a short fade at each edge, so the sample starts and ends without a click. Each cut is saved as its own file.

Is my song uploaded?

No. Decoding, separation, playback and export all happen in your browser tab. The only download the app makes is the model itself, from Hugging Face, once. You can confirm it yourself: open your browser's developer tools, watch the Network tab, and run a separation.

A note on rights

Separating a song for practice, study, remixing at home or a karaoke night is one thing; releasing something built on someone else's recording is another. The stems are still the original recording, so clear the rights before you publish them.