Add to Chrome

Stem Separation

Split a song into vocals, drums, bass and other parts with an AI model — processing stays in your browser.

Audio tool

Split a song into four stems

Runs in your browser

Drop a song here

or click to choose from your computer (up to 5 minutes)

MP3 · WAV · FLAC · AAC · M4A · OGG

✓ Local processing in your browser · Files are not uploaded to our servers

What is Stem Separation?

Stem Separation splits a finished song into its separate parts, called stems: vocals, drums, bass and other (guitars, keys, synths and effects). You get four audio files that you can listen to, mute or reuse in another project.

Use it to make a karaoke track, to practice drums or bass over the real recording, to pull a vocal for a cover or remix, or to fix one instrument in a mix. The separation runs in your browser with a neural network, so the song never leaves your device.

Stems are an estimate, not a perfect split. The better the source file, the cleaner the stems. A dense mix or a heavily compressed MP3 leaves more traces of one part in another.

Stem separation is a different job from the Vocal Remover. The Vocal Remover subtracts the centered part of a stereo mix and gives you one instrumental. Stem Separation uses a neural network that has learned what drums, bass and vocals sound like, so it returns several separate tracks that you can mix again, mute or export one by one.

Typical uses: a karaoke track with the voice removed, a drum or bass part to play along with, a clean vocal for a cover or remix, or one instrument pulled out of a demo to study its arrangement. Because you get the parts separately, you can also rebuild the song with, say, a louder bass and a quieter drum part in your DAW.

Keep your expectations realistic. Separation is strongest on modern, well-recorded pop, rock and electronic music with a clear vocal. Very dense mixes, live recordings with a lot of room sound, and heavily distorted guitars are harder, and some of their energy can end up in the wrong stem.

How the stem separation works

The tool uses Spleeter 4stems, an open-source model from Deezer. It has four small networks, one for each stem, and they look at the song in short time-frequency slices.

01

Read the song

The file is decoded to stereo 44.1 kHz and cut into overlapping 4096-sample windows. Each window is turned into a frequency map, a spectrogram.

02

Estimate each stem

For every part of the spectrogram, each network predicts how much of the sound belongs to vocals, drums, bass or other.

03

Share out the sound

The four estimates are compared cell by cell, so each part of the spectrogram is divided between the stems and the shares always add up to the original.

04

Rebuild the audio

The shares are applied to the original signal and converted back to sound. The four stems add up to the song you loaded.

Tips for better stems

  • Start from the best file you have. A WAV or 320 kbps MP3 gives cleaner stems than a 128 kbps file.
  • Expect small bleed. Cymbals can show up in vocals, and vocals can leak into other. This happens with every separator, so listen to the stem in context before you use it.
  • Trim first. Files longer than 5 minutes are rejected. Use the Audio Trimmer to cut the part you need.
  • Mind the device. Separation works best on a desktop. Phones can run out of memory on longer songs.

Frequently asked questions

What are stems?

Stems are the separate parts that were mixed together to make a song: usually vocals, drums, bass, and everything else (guitars, keys, synths and effects), which this tool calls Other. Each stem is a full-length audio file you can play, mute or reuse.

How is this different from the Vocal Remover?

The Vocal Remover subtracts centered sound from a stereo mix. It is fast, but it gives you one instrumental and removes anything in the center. Stem Separation uses a neural network trained to recognize each part, so you get four stems. Vocals usually come out cleaner, but processing takes longer.

How long does processing take?

It depends on the length of the song and on your computer. A progress bar shows how far along it is. The model (about 80 MB in total) is downloaded once and then reused from the browser cache.

Which model does it use?

Spleeter 4stems, an open-source model published by Deezer under the Apache 2.0 license. It was converted to ONNX so it runs in your browser with ONNX Runtime Web.

Are my songs uploaded anywhere?

No. The song is decoded, separated and encoded on your device. The stems are downloaded straight from the page, and nothing is sent to Sound Booster servers.

Why is there a 5-minute limit?

The whole song is held in memory while it is separated, and longer files can make the browser slow or unstable. Trim longer tracks first with the Audio Trimmer, then separate the part you need.

Which format should I download?

MP3 at 320 or 256 kbps is fine for practice, karaoke and demos. Choose WAV if you plan to edit the stems in a DAW, so the audio is not compressed twice.

Can I publish the stems?

The tool does not change the rights to the original song. Stems are derived from that song, so you can use them only as far as you have the right to use the original recording.

Related tools

Work with the stems further in the browser.