This vocal remover and stem splitter makes karaoke tracks, a cappellas and separate drum, bass and instrument stems from your songs, entirely in your browser. Choose the AI engine and it downloads Demucs, Meta's music separation model, once from Hugging Face into your browser, then splits any song — stereo or mono, with reverb and backing vocals — into drums, bass, other and vocals, using your GPU where available; you can delete the model at any time. Or choose the instant engine, which needs no download and uses centre-channel cancellation: in almost every studio mix the lead vocal sits dead centre, so subtracting one stereo channel from the other removes it, while guitars, keys and backing parts panned to the sides remain. Plain cancellation also wipes out the centred bass and kick drum, so this tool puts the low end back through a filter whose cut-off you can adjust. Drop a single song or a folder of them, preview each result in the page, and download 16-bit WAV files one by one or as a ZIP. A rough vocals-only output is there too. Either way it is free, works on a whole folder of songs, lets you preview each result, downloads 16-bit WAV files or a ZIP, and your music never leaves your device.
| Part of the mix | Usual position | In the instrumental |
|---|---|---|
| Lead vocal | Centre | Removed |
| Bass guitar and kick | Centre | Kept (restored below the cut-off) |
| Snare | Centre | Mostly removed |
| Guitars and keys | Panned left or right | Kept |
| Backing vocals | Often panned wide | Kept |
| Vocal reverb | Spread across stereo | Partly kept |
The AI engine (Demucs) has learned what voices and instruments sound like, so it separates them cleanly even with reverb, backing vocals or a mono file, and can output four stems. It needs a one-time download of about 172 MB and takes about a minute per song on a GPU. The instant engine needs no download and works in seconds, but only removes what is mixed dead centre.
Yes. Use Clear downloaded models; after confirming, it is removed from your browser and would be downloaded again next time.
Lead vocals are usually mixed equally into both stereo channels. Subtracting one channel from the other cancels anything in the centre, which removes the voice; the bass is then restored from the low frequencies.
Reverb and echo on the voice, and backing vocals panned to the sides, are not in the exact centre, so they survive. Live recordings and some modern mixes cancel less cleanly.
The file is mono or nearly mono, so both channels are the same and almost everything cancels. Use a stereo version of the song.
No. Decoding, processing and WAV encoding happen in your browser.
16-bit WAV at the song's own sample rate, which any player or editor can open.
Songs are decoded by the browser at their own sample rate. Lead vocals are almost always mixed in the centre, equally loud in both channels, so the side signal (left − right) ÷ 2 cancels them. That also cancels centred bass and kick drum, so the low end of the mid signal (left + right) ÷ 2 is restored through a fourth-order low-pass filter (two cascaded RBJ biquads) at an adjustable cut-off, 120 Hz by default. The vocals option does the reverse: the mid signal high-passed above the cut-off, a rough centre extraction. A stereo check measures the correlation between channels and the side-to-mid level, and warns when a file is effectively mono. The AI engine runs HTDemucs, Meta's hybrid transformer source-separation model (MIT), exported to ONNX with the STFT outside the graph and executed with ONNX Runtime Web on WebGPU or WebAssembly in a Web Worker. The model (about 172 MB) downloads only when the visitor asks, straight from Hugging Face into the browser's Cache Storage, and can be removed from the Downloaded models panel. Audio is resampled to 44.1 kHz and cut into 7.8-second segments with 25% overlap; each segment's waveform and its 4096-point STFT (hop 1024) go in, the model returns a time-domain branch and a complex spectrogram branch per source, the spectrogram is inverted and the two are summed, and segments are overlap-added with tapered weights into drums, bass, other and vocals. Instrumental is drums + bass + other. Output can be peak-normalised and is written as a 16-bit PCM WAV.