Audio
Vocal Remover
Split any song into vocals and instrumental — or full drums/bass/other stems — using AI, entirely in your browser. Your music never leaves your device.
Drop your song here
or click to browse · MP3, WAV, M4A, AAC, OGG, FLAC
This vocal remover runs Demucs — an MIT-licensed AI music source separation model from Meta AI — directly in your browser using ONNX Runtime Web (WebAssembly, with WebGPU acceleration where your browser supports it). There is no server, no account, and no upload: the model itself is downloaded once and cached by your browser, and every separation after that runs entirely offline on your machine. Nothing about your audio is ever sent over the network — you can confirm this yourself by watching the Network tab in your browser's developer tools while a separation runs.
Choose 2-stem (Vocals + Instrumental) for the most common use cases — karaoke tracks, acapellas, or a clean instrumental to sing or play over — or 4-stem for Vocals, Drums, Bass, and Other separately. Quality is very good on most modern mixes; extremely dense or heavily-mastered tracks can leave faint bleed between stems, which is a property of AI source separation in general.
How to Use
- 1
Drop a song onto the uploader or click to browse. Supported formats: MP3, WAV, M4A, AAC, OGG, FLAC.
- 2
Choose 2-stem (Vocals + Instrumental, fewer files) or 4-stem (Vocals, Drums, Bass, Other), and your preferred output format (MP3 or WAV).
- 3
Click "Separate". On first use, the AI model is downloaded and cached by your browser — after that, it starts immediately. Separation itself typically takes a few minutes on desktop, with real progress shown throughout.
- 4
Preview each stem in the built-in audio player, then download the ones you need.
Frequently Asked Questions
Is my music uploaded to a server?
No. Separation runs entirely in your browser via ONNX Runtime Web. Your audio never leaves your device — verifiable in your browser's Network tab during a run.
How long does it take?
A few minutes for a typical song on desktop, depending on track length and hardware. The tool shows real segment-by-segment progress plus elapsed and estimated-remaining time, not a fake spinner.
2-stem vs 4-stem — which is faster?
Neither — the AI model always analyzes every instrument in one pass, so processing time is identical. 2-stem (Vocals + Instrumental) just means fewer output files to encode and download, which is why it's the default.
Will it be perfect?
Very good, but not perfect. Dense or heavily-mastered mixes can leave faint bleed between stems — a trace of one instrument in another. This is inherent to AI source separation, not specific to this tool.
Does this work on mobile?
It can, but this is a heavy, desktop-oriented tool. On mobile or lower-memory devices you'll see a clear warning before starting, and very long tracks are blocked outright to avoid the tab crashing. For best results, use a desktop browser.