Separate a song into vocals and instrumental. Listen to both, then download.
Runs locally in your browser
Or choose a file from your device.
MP3, WAV, M4A, FLAC, OGG, AIFF, WMA & moreYour audio stays on this device · Up to 128 MB per fileDrag and drop from your device.
AI separates the voice and music.
Listen to the result and download.
Open MP3, WAV, M4A, AAC, FLAC, OGG, OPUS, AIFF, WMA, AMR, AC3 and other audio formats supported by the included FFmpeg decoder. You can also extract audio from video files. Support depends on the codec inside the file; damaged or protected files may not open.
Export MP3, WAV, M4A, FLAC, OGG, or M4R ringtones. The converter downloads on first use. Files stay on your device. Refreshing this page clears the current session.
Use arrow keys on the selection handles for precise adjustments. Shift + arrow moves by one second. Press Escape to stop playback.
Complete tool guide
Separate a song into two tracks: an instrumental without the lead vocals, and an isolated vocal (acapella). The separation uses the open-source Hybrid Transformer Demucs model running on your device, so the song is never uploaded. Preview both tracks, then download them as WAV files.
Your input is processed in the current browser tab and is not sent to the application server.
Choose or drag in a song up to 10 minutes long. Use the trimmer first for longer files or to process only one section.
Select Separate vocals. The first time, the tool downloads an AI model of about 172 MB; later runs reuse the cached copy when your browser allows it.
Wait while the song is processed in segments. Progress is shown, and you can cancel at any time.
Listen to the Instrumental and Vocals results, then download either or both WAV files.
The model splits music into drums, bass, other instruments, and vocals. The instrumental is the sum of the drums, bass, and other stems.
The vocal stem contains the lead voice and often backing vocals and harmonies. Reverb and delay effects on the voice are usually included with it.
Both tracks are 44.1 kHz stereo 16-bit WAV files, about 10 MB per minute each. If the separated audio would clip, both stems are lowered by the same amount.
Start with: song.mp3, 3:30 long
song-instrumental.wav · 3:30.00 · about 35 MB
song-vocals.wav · 3:30.00 · about 35 MBStereo 16-bit WAV at 44.1 kHz uses 176,400 bytes per second, so 210 seconds is about 37 million bytes, or 35.3 MB, per stem.
Start with: track.flac, 5:10 long, chorus from 1:20 to 2:20
track-cut-vocals.wav · 1:00.00 · about 10 MBProcessing one minute instead of the whole song finishes roughly five times faster and uses far less memory.
Hybrid Transformer Demucs is a neural network trained to predict each instrument group from a mixed song. It runs through ONNX Runtime in a background worker using WebAssembly, so the page stays responsive while it works.
Separation is computationally heavy. Depending on your device, it can take from about the length of the song to several times longer. A desktop or laptop with plenty of free memory gives the best results; phones may run out of memory.
Modern separation is good but not perfect. Faint vocal traces can remain in the instrumental, and cymbals or synths can leak into the vocal, especially with heavy effects or dense mixes.
Yes. There is no sign-up or per-song limit. The song is processed on your device; the only download is the AI model, which comes from a public model host the first time you use the tool.
That is the size of the Demucs neural network. Running it locally is what keeps your music on your device. The browser usually caches it, so later separations start much faster.
It depends on the song length and your device. A three-minute song typically takes a few minutes on a modern laptop. Processing stops automatically after 20 minutes to protect slower devices.
Yes. Download the Vocals track: it is the voice with the music removed. This works for songs and can help with speech recorded over music.
Separating a song does not change who owns it. Publishing a stem generally requires permission from the rights holders; use the results for personal practice unless you have that permission.
Keep working
Continue with another focused tool from the same collection.