Loading tool…
Need a karaoke track, an a cappella or a practice backing track? Drop in a song and an MDX-Net source separation model splits it into a vocal stem and an instrumental stem on your own computer. You can solo and balance both stems before downloading them.
Processing locally on your device
Overview
In a finished stereo mix the voice and the instruments share the same frequencies and the same two channels, so no filter can pull them apart cleanly. The old karaoke trick subtracts one channel from the other, which removes whatever is panned dead centre. That usually includes the lead vocal, but also the bass, kick and snare, and it leaves the vocal reverb behind. Neural source separation works differently. The model has learned what singing sounds like and estimates, for every moment and every frequency, how much of the mix belongs to the voice.
This tool runs UVR MDX-Net models, the KUIELAB architecture that placed in the 2021 Music Demixing Challenge and is used by the open-source Ultimate Vocal Remover. They're published as ONNX files under the MIT licence. Three are offered. Instrumental (Inst HQ 3, 67 MB) gives the cleanest backing track. Vocals (Voc FT, 67 MB) is tuned for a cappellas. Karaoke (Kara 2, 53 MB) removes the lead singer but keeps the backing vocals. The model is downloaded once and then cached by your browser. The song is resampled to 44.1 kHz stereo and cut into chunks of about six seconds, and each chunk is turned into a spectrogram in JavaScript (a 6,144-point STFT for the default model). ONNX Runtime then predicts the instrumental for that chunk. The vocal stem is whatever the model takes out of the mix, so the two stems always add back up to the original. Neighbouring chunks overlap and are cross-faded, so there are no clicks at the seams.
This is heavy work for a browser. With a WebGPU graphics card the network runs on the GPU, which is typically much faster. Without one it runs on a single CPU core in WebAssembly, where each minute of audio can take several minutes. To check a model quickly, select a 30-second section on the waveform and separate just that part. There is also an Instant mode that needs no download. It is a frequency-domain centre-channel extractor that keeps the bass and anything panned off-centre, but like any non-AI method it can't remove reverb or a vocal that isn't centred, and it does nothing on mono recordings.
Once the stems are ready, the mixer plays both in perfect sync. Each stem has its own volume, mute and solo controls, so you can turn the vocal down to 20% for a sing-along guide instead of removing it completely. You can export the vocals, the instrumental, both as a ZIP, or the mix exactly as you've set it, in MP3, M4A, WAV, FLAC, OGG or Opus.
| Mode | Download | Best for | Limitations |
|---|---|---|---|
| Instrumental (Inst HQ 3) | 67 MB | Clean backing and karaoke tracks | Slowest on CPU |
| Vocals (Voc FT) | 67 MB | A cappellas, vocal study | Instrumental may keep more vocal bleed |
| Karaoke (Kara 2) | 53 MB | Removing only the lead singer | Backing vocals stay in the instrumental |
| Instant (centre cancel) | None | Quick tests on stereo mixes | Leaves reverb; useless on mono |
Drop a song (MP3, WAV, M4A, FLAC, OGG) or a music video into the tool.
Pick a model: Instrumental, Vocals, Karaoke, or Instant for a no-download centre cancel.
Optionally drag across the waveform to separate only a section first.
Click Separate stems; the first run downloads the model once.
Play the stems in the mixer, using solo, mute and the volume sliders to check them.
Choose a format and download the vocals, the instrumental, both, or your custom mix.
Use cases
Remove the lead vocal from a favourite song for a party or home practice. The Karaoke model keeps the backing harmonies, which makes the result feel closer to an official karaoke version.
Rather than deleting the vocal, turn it down to around 20% in the mixer and download that mix. You still hear the melody as a guide while your own voice carries the song.
Pull the vocal out of a track to use in a remix, mashup or DJ edit. Check that you have the rights to use the original recording before publishing anything built from it.
Solo the instrumental to work out chords and bass lines without the vocal getting in the way, or solo the vocal to learn harmonies and phrasing.
Pro tips
Answers to common questions about the vocal remover.
Instrumental (Inst HQ 3) gives the cleanest backing track and is the best default. Vocals (Voc FT) is tuned to isolate the singing for an a cappella. Karaoke (Kara 2) removes only the lead singer and leaves backing vocals and harmonies in the instrumental.
The network processes a large spectrogram for every six seconds of audio. On a WebGPU graphics card that's quick, but without one it runs on a single CPU core in WebAssembly, where each minute of music can take several minutes. Separating a shorter section is the fastest way to try it.
Close, but rarely perfect. Expect occasional faint vocal remnants, softened cymbals or a slightly watery sound where the voice and an instrument play the same notes. Cleanly produced modern mixes give the best results.
It compares the left and right channels for every frequency band and moves whatever is identical in both, meaning centre-panned, into the vocal stem. Bass and low frequencies are left alone. It is instant and needs no download, but reverb, doubled vocals and centred instruments aren't separated correctly, and mono files can't be split this way at all.
Not in this tool. It produces two stems, vocals and instrumental. Four-stem models such as Demucs are several times larger and too slow to run on a typical CPU in the browser.
No. ToolEnhancer processes files on your own device inside the browser. They are never sent to, stored on or seen by our servers.
Yes. The vocal remover is free to use with no sign-up, no watermarks and no daily limits.
Transcribe audio and video privately.
Natural voices, generated on-device.
Clean up voice recordings.
Cut “um” and “uh” automatically.
Translate subtitles on-device.
Remove image backgrounds in one click.