All tools Separate parts
Vocal Remover
Split a song into vocals and instrumental with an AI model that runs in your browser, then download either one. The song never leaves this device; only the model downloads, once.
Drop an audio file here
MP3, WAV, M4A, FLAC, OGG or WebM
Your file stays on this device. Nothing is uploaded.
Reading your file
0:00 / 0:00
Splits the song into two parts: the voice (acapella) and everything else (instrumental, for karaoke).
- The separation model downloads once (90.4 MB) and stays saved in this browser.
- Checking this device…
This device has too little memory for AI separation
The separation model needs about 2 GB of working memory, and this browser reports less than 4 GB in total. Use a computer, or a phone or tablet with more memory.
Too long to separate without WebGPU on this device
This will take a while on this device
The saved model was deleted from this browser.
The graphics processor stopped, so the tool continues on the processor. This is slower; the parts already done are kept.
Vocals and instrumental are ready
- Vocals
- Instrumental
Separated parts are 44.1 kHz stereo, the rate the model works at. Every file goes through a limiter, so peaks stay at or below −1 dBFS.
Want the instrumental in another key? Try Pitch Changer
This isn’t an audio file
Choose an MP3, WAV, M4A, FLAC, OGG or WebM file and try again.
How to remove vocals from a song
Choose a song
Drop an MP3, WAV, M4A, FLAC, OGG or WebM file. It opens in this browser tab and is not uploaded.
Press Remove vocals
The first time, the browser downloads the separation model (90.4 MB) and saves it. The song is then split piece by piece, with the time left shown.
Listen, then download
Solo or mute the vocals and the instrumental while it plays. Download either part as MP3 or WAV, or both in one ZIP.
What happens on your device
The tool runs Demucs, an open-source music separation network, inside this tab. It uses the graphics processor through WebGPU when the browser offers it, otherwise the main processor. The song is cut into 7.8-second pieces that overlap, each piece is separated, and the pieces are joined back. The instrumental is the song minus the separated vocals, so the two parts add up to the original.
Nothing is uploaded: no server receives your audio. This browser remembers the settings you pick, never the song.
Limits on your device
Everything runs in your browser, so the longest song depends on your device’s memory. The tool checks the length before it starts and tells you if a file is too long.
| Device | Longest song | With little memory |
|---|---|---|
| Laptop or desktop | 15 min | 6 min (4 GB of memory or less) |
| Phone or tablet | 6 min | 4 min (4 GB of memory or less) |
Separation needs much more memory than the other tools: every part is a full copy of the song, and the model itself takes memory while it runs. That is why the longest song here is shorter than for the effects. The model also needs about 2 GB of working memory while it runs, so a device with less than 4 GB of memory is told before anything downloads. Without WebGPU the model runs inside the tab and is much slower, so the longest song on that path is 8 minutes on a computer and 4 on a phone.
- Opens
- MP3, WAV, M4A, FLAC, OGG and WebM
- Downloads
- Each part as MP3 (128–320 kbps) or WAV (16-bit) at 44.1 kHz, or all parts in one ZIP
- Speed we measured
- Measured on a Windows PC with an RTX 3050 graphics card, in Chrome and Edge: a 4-minute song separates in about 31 seconds with WebGPU, plus a few seconds to load the model. Without WebGPU the same song takes about 12 minutes. Phones and computers without a graphics card are slower.
Separation speed, memory and quality by device: our full measurements
Questions
Is my song uploaded to a server?
No. Only the separation model is downloaded from this site, once. Your song is decoded and separated in this browser tab and never leaves the device.
How good is the separation?
It uses Demucs (HTDemucs fine-tuned for vocals), an open-source separation model. On our test song, made in code so every part is known, the vocal part scored 12.4 dB SDR, the usual measure for separation quality. Clear lead vocals separate well; heavy reverb, backing choirs and vocals buried in distorted guitars can leave traces in either part. Listen with Solo before you download.
Why does it download a model, and how big is it?
The network that recognizes the voice is a file of about 90.4 MB. It downloads the first time and stays saved in this browser, so the next song starts at once. You can delete it with “Delete the saved model”.
How long does it take?
With WebGPU (Chrome or Edge on a computer with a graphics card) a 4-minute song takes well under a minute. Without it, the processor does the work on one core, which is many times slower. The tool shows an estimate before you start and the time left while it runs.
Can I make a karaoke track in another key?
Yes. Download the instrumental, then open it in Pitch Changer to move it up or down in semitones without changing its speed.