Free AI vocal remover and stem splitter

How to remove vocals from a song

1. Choose the song

   Drop an MP3, WAV, M4A, FLAC or OGG file into the box, or choose a video. The browser reads it; nothing is uploaded.

2. Press Remove vocals

   The first time, the AI model downloads once (171 MB). The song is then separated in parts of about eight seconds, and the progress bar shows how far along it is.

3. Listen and adjust

   Instrumental plays the song without the voice. Vocals only gives you the acapella. Switch single stems on or off and change their volume to make your own mix.

4. Download

   Save the instrumental, the vocals or any single stem as a WAV file in CD quality (44.1 kHz, 16-bit, stereo).

Four stems from one song

The Vocal Remover does not just cut out the middle of the stereo picture like old karaoke tricks. An AI model listens to the whole song and rebuilds it as four separate tracks, called stems:

• Vocals – lead and backing voices, the acapella
• Drums – kick, snare, cymbals and other percussion
• Bass – bass guitar, synth bass and low notes
• Other instruments – guitars, piano, strings, synths and everything else

Played together, the four stems sound like the original. The instrumental is simply drums, bass and other instruments without the vocals, which is what most people mean when they search for a vocal remover or a karaoke version of a song.

How the AI separates a song

The separation uses Hybrid Transformer Demucs (htdemucs), an open-source model from Meta’s research team, trained on thousands of songs to tell voices and instruments apart. It looks at the sound twice at the same time: as a waveform and as a spectrogram, the picture of which frequencies sound when. A transformer connects both views, so it can follow a voice even when a guitar plays the same notes.

The song is processed in overlapping parts of about eight seconds, and the overlaps are blended smoothly so you hear no seams. All of this runs on your device with ONNX Runtime Web, on the graphics chip when the browser supports WebGPU and on the processor otherwise.

Tips for the cleanest result

• Start from the best file you have. A WAV, FLAC or high-bitrate MP3 separates more cleanly than a low-quality copy.
• Studio recordings work best. In live recordings the crowd and the room echo mix with the voice and some of it stays in the instrumental.
• Heavy reverb and effects on the voice can leave a faint trail in the instrumental. Lower the vocal stem instead of muting it, or keep a little of it for a guide vocal.
• Backing choirs that sound like an instrument sometimes land in Other instruments. Listen to the vocals stem and the other stem to find them.
• For karaoke, try the instrumental with the vocals stem at 10 to 20 percent: a quiet guide voice helps you keep the melody.

Mix, loop, change key and speed

After the separation the Vocal Remover works like a small studio: every stem has a waveform, mute (M), solo (S) and a volume slider, and presets switch between Instrumental, Karaoke with a quiet guide voice, Vocals only, No drums, No bass and All stems.

• Key – up to six semitones up or down, tempo unchanged.
• Speed – 50 to 150 percent, pitch unchanged.
• Loop – repeat one part to practise it.
• BPM and key – estimated from the drums and the harmonic stems.
• Downloads – single stems, instrumental, vocals, your mix or all stems as a ZIP, with the changed key and speed.

High precision separates every part twice and averages the results; in our test that was about 0.6 dB cleaner.

Privacy and music rights

Your song is never sent to a server: it is decoded, separated and saved by your browser, so it is private by design. Removing vocals for your own karaoke night, practice or study is fine. If you want to publish a remix, a cover backing track or a video that uses the stems, make sure you have the rights to the original recording, as the song still belongs to its owners.