At a glance
- Accepts
- MP3, WAV, M4A, AAC, OGG, Opus, FLAC, and the audio track of a video
- Outputs
- Four 44.1 kHz stereo tracks (vocals, drums, bass, other) as MP3 (192 kbps) or WAV, one at a time or together in a ZIP; the four add back up to the original exactly
- Where it runs
- Entirely in your browser, on your device
- What leaves your device
- Nothing - the file never leaves your device
- Limits
- Downloads a 158 MB model once; about 1-3 minutes per song with WebGPU, often 10+ minutes on a processor alone; phones and tablets can run out of memory; dense or heavily effected mixes leave some bleed between stems, and reverb on a voice often lands in other
- Account
- Not required - there is no sign-up
- Price
- Free, with no watermark and no usage cap
- Made by
- Visrut
- Last updated
- 23 October 2026
Split a song into four stems: vocals, drums, bass and everything else. Listen to each one, then download them as MP3 or WAV, one at a time or all four in a ZIP. The AI model runs in your browser, so the song never leaves your device.
How to use it
- Drop in a song, or a video to use its soundtrack.
- Press Split into 4 stems. The first time, the AI model downloads and is saved for next time.
- Play each stem, then download the ones you need.
How it works
The tool uses HT-Demucs, a hybrid transformer model published by Meta AI researchers that looks at both the waveform and the frequencies of the music. The song is cut into overlapping 7.8-second pieces, each piece goes through the model, and the pieces are cross-faded back together so there are no clicks at the joins. The model predicts the vocals, drums and bass; the fourth stem is everything that is left, so the four always add up to the original. It runs on your graphics chip with WebGPU when it can, and on the processor otherwise.
Only need two tracks?
For a karaoke version or an acapella, Remove Vocals is quicker and has a small model that works on phones. To sing the instrumental in a different key, take it to Change Song Key & Tempo, and the BPM & Key Finder tells you the tempo and key for a remix.
Frequently asked questions
What are stems?
Stems are the separate parts a song is made of. This tool gives you four: vocals, drums, bass, and everything else (guitars, piano, synths, strings) in a fourth track. Producers use them for remixes and mashups, DJs for acapellas and instrumentals, and musicians to practise along with a song with their own part taken out.
How good is the separation?
It uses HT-Demucs, the open-source model from Meta's research team that many paid stem splitters are built on. On the standard MUSDB18-HQ benchmark it scores about 8.8 dB for vocals and 9.5 dB for drums, which in practice means clean drums and vocals on most pop, rock and hip-hop, a usable bass line, and some bleed between tracks on dense or heavily effected mixes. Reverb on a voice often stays partly in the other track.
Do the four stems add back up to the song?
Yes, exactly. The vocals, drums and bass come from the model, and the other track is whatever is left over, so if you play all four together you hear the original song with nothing lost.
Why does it need a 158 MB download?
The AI model runs on your own computer, so it has to be downloaded once. After that it is saved in your browser and the next song starts straight away. Nothing about your song is sent anywhere: the model comes to the file, not the other way round.
How long does it take?
On a computer with a graphics chip that supports WebGPU (most recent laptops in Chrome or Edge), a four-minute song takes about one to three minutes. Without one, the model runs on the processor and can take ten minutes or more. The page shows an estimate before you start and a countdown while it runs.
Does it work on a phone?
It can, but the model needs a lot of memory, and phones and tablets often run out part-way through. A laptop or desktop works best. If you only need the vocals or the instrumental, Remove Vocals has a smaller, faster model that works well on phones.
Is my song uploaded?
No. The song is decoded and separated inside your browser tab, and the stems are made there too. The only download is the AI model itself, from Hugging Face, the first time.