How to separate vocals from music and get an acapella or a karaoke track
Split any song into vocals and instrumental with an AI model that runs in your browser: steps, how long it takes and how to get cleaner stems. No upload.
Updated
Short answer: open the Vocal remover, drop the song and press Separate vocals. You get two files: Vocals (acapella) and Instrumental (karaoke). The AI model runs on your own computer or phone, so the song is never uploaded. The first run downloads the model once (about 172 MB); after that it works offline.
Steps
- Open the Vocal remover in Chrome or Edge on a computer if you can. They use your graphics card through WebGPU, which is much faster.
- Drop a song onto the page: MP3, WAV, FLAC, M4A, or a music video. A video’s soundtrack is used directly, no need to convert it first.
- Press Separate vocals. You see the model download on the first run, then Separating… with the time left.
- Download Vocals (acapella), Instrumental (karaoke), or both. For remixing, take All 4 stems as a ZIP: vocals, drums, bass and other instruments.
The files are MP3 by default; you can pick another format before downloading.
How long it takes
With a graphics card, a typical song takes a minute or two once the model is loaded; the page shows an estimate after the first seconds. Without WebGPU (Firefox, many phones, old laptops) the page warns you and runs on the processor, roughly 8 minutes of work per minute of music. It still works, it’s just slower, so leave the tab open.
If a long file runs out of memory, cut it first with the Audio cutter and separate the part you need.
What the AI does
The tool uses Hybrid Transformer Demucs (htdemucs) from Meta AI Research, one of the best open models for music separation, released under the MIT licence. It learned from thousands of songs what voices, drums and bass sound like, and splits the mix into four stems. “Vocals” is the voice stem; “Instrumental” is the other three added back together.
Older “karaoke” tricks just cancel whatever sits in the middle of the stereo mix. That also removes bass and kick drum and leaves reverb behind. A trained model keeps the instruments intact.
How to get cleaner results
- Start from the best source. A WAV, FLAC or high-bitrate MP3 separates better than a heavily compressed clip ripped from a messaging app.
- Expect some bleed. Backing vocals, choirs and heavy reverb can stay in the instrumental, and loud cymbals may leave a faint trace in the acapella. That’s normal for every separator.
- Fix the level afterwards. Stems are often quieter than the full song. Normalize audio brings them to a standard loudness.
- Speech, not songs? For a voice recording with traffic, wind or a fan behind it, Clean up a voice recording is the better tool.
What people use it for
Karaoke and singing practice, learning a bass or drum part, acapellas for remixes and DJ sets, sampling, transcribing lyrics, or a backing track for a school performance. Remember that the song still belongs to its rights holders: separated stems of someone else’s track are for personal use unless you have a licence.
FAQ
Is my song uploaded?
No. The AI model comes to your device, not the other way round. The file never leaves the browser tab.
Does it work on a phone?
Yes, in a current browser. Most phones run it on the processor, so a song takes several minutes, and the 172 MB model download is better done on Wi-Fi.
Is it free? Is there a length limit?
It’s free. The practical limit is your device’s memory: full songs work fine, very long DJ mixes may need to be cut into parts.
Can I remove the music and keep only the voice from a video?
Yes. Drop the video file itself and download Vocals (acapella). It works for songs and also for speech over background music.