At a glance
- Accepts
- MP3, WAV, M4A, AAC, OGG, Opus, FLAC, and the audio track of a video
- Outputs
- Two 44.1 kHz stereo WAV files: the instrumental (karaoke track) and the isolated vocals
- Where it runs
- Entirely in your browser, on your device
- What leaves your device
- Nothing - the file never leaves your device
- Limits
- The AI mode downloads a model once: Fast quality (Spleeter, 38 MB) is quick on any device but leaves more bleed between vocals and instrumental; Best quality (HT-Demucs, 158 MB) is cleaner but slow without WebGPU and can run out of memory on phones; reverb and backing vocals can leave a faint trace of the singer; instant methods can affect centred instruments
- Account
- Not required - there is no sign-up
- Price
- Free, with no watermark and no usage cap
- Made by
- Visrut
- Last updated
- 14 October 2026
This vocal remover turns a finished song into an instrumental you can sing over. It runs an open-source separation model directly in your browser, so the song is never uploaded anywhere: Spleeter for Fast quality, or HT-Demucs fine-tuned for vocals for Best quality. You get two files: the instrumental, which is your karaoke track, and the isolated vocals as a bonus.
How to use it
- Drop in a song, or a music video to use its audio track.
- Keep the AI vocal remover selected and pick Fast or Best quality. The first run downloads the model (38 MB for Fast, 158 MB for Best), then it is cached for every song after.
- Press remove vocals, play the instrumental in the page, and download it as a WAV. The isolated vocals are right below it if you want them.
Why the old karaoke trick is not enough
The traditional way to strip vocals is to subtract one stereo channel from the other. It works because singers are mixed to the centre, but bass, kick drum and snare are mixed there too, so they vanish with the voice and the result sounds thin and hollow. This page's instant mode extracts the shared sound in the vocal frequency range, preserving low bass and stereo width. It can still affect centred instruments; mono recordings use background subtraction instead. The AI model learned what a voice looks like in a spectrogram, so it can separate vocals more selectively.
What it will not do
Separation is an estimate, not an undo button. Heavy reverb, doubled vocals and harmonies can leave a faint trace of the singer in the instrumental, and anything the voice was masking in the original mix cannot be recovered. On a CPU the model can take several minutes for a full song; a browser with WebGPU is much faster.
Sources
Spleeter's vocal SDR on MUSDB18 (6.55 dB with soft masking, 6.86 dB with Wiener filtering, for its 4-stem model): Hennequin et al., JOSS 2020. HT-Demucs FT median vocal SDR on MUSDB18-HQ (9.19 dB): the model card of the ONNX export used here; the model itself is from Rouard, Massa and Défossez, 2022. The browser version of Spleeter is the sherpa-onnx conversion of Deezer's 2-stem model (MIT licence). Our own test numbers use the simple SDR of the MDX challenge, so they are not directly comparable with the published ones. Checked October 2026.
Frequently asked questions
Is my song uploaded to a server?
No. The song is decoded and separated inside your browser tab. The only network traffic is the separation model downloading to you the first time you use the AI mode - your file never leaves your device.
How do I make a karaoke version of a song?
Drop the song in, keep the AI model selected, and press the button. When it finishes, the instrumental is the karaoke track: play it in the page to check it, then download it as a WAV.
What is the difference between Fast and Best quality?
They are two different AI models. Fast is Spleeter 2-stems from Deezer, a 38 MB download that splits a 3-minute song in seconds on a recent computer and is light enough for phones. Best is HT-Demucs fine-tuned for vocals, a 158 MB download that is cleaner but needs much more memory and is slow without WebGPU. Fast leaves more of the instruments in the vocals and more of the voice in the instrumental. On the MUSDB18 benchmark the published vocal scores are about 6.6 dB SDR for Spleeter (its 4-stem model, soft masking) and 9.19 dB for HT-Demucs FT (higher is cleaner; sources linked below). In our own test on a 3-minute song with known stems, Fast scored 5.8 dB and Best 8.7 dB on the vocals. The page picks Fast where Best would be slow or run out of memory, and you can run Best afterwards on the same song with one click.
Why does the AI mode download a model?
The separation runs on your device, so the model's weights have to come to your browser: 38 MB for Fast quality, 158 MB for Best. They are cached after the first run, so the next song starts straight away.
What does the instant mode do, and when does it work?
It extracts sound shared by both stereo channels in the vocal frequency range, reducing centred vocals while preserving low bass and stereo width. Centred instruments and backing vocals can also be affected. Mono recordings use background subtraction instead, which works best with a steady musical bed under speech. Neither method can separate every song as cleanly as the AI model.
Will the instrumental sound perfect?
Not always. Reverb and backing vocals often leave a faint ghost of the singer, and a voice buried in a dense mix comes back with some artefacts in the instruments around it. For singing along it is usually more than good enough; for a release, listen carefully first.
Can I get the vocals on their own too?
Yes. Every run produces both files: the instrumental, and the isolated vocals (an a cappella). The instrumental is your original minus the vocals, so layering the two back together gives you the original song.
Can I remove the piano, drums, bass or guitar instead?
Not with this tool. It splits a song into two parts: the vocals, and everything else. Piano, guitar, drums and bass all stay together in the instrumental, so it can make a version without vocals but not one without piano. The AI model does make rough drum, bass and 'other' estimates internally, but this version of it is tuned for vocals, and in our tests those parts were not clean enough to offer: on a mix of solo piano and a voice, with no drums at all, its drum part still picked up a noticeable share of the piano. The one open model small enough for a browser that has a piano part, Demucs htdemucs_6s, is described by its own authors as having a lot of bleeding and artifacts on piano. In our test its 'no piano' track kept much of the piano, so we have not added it.
What formats can I use?
Anything your browser can decode: MP3, WAV, M4A, AAC, OGG, Opus, FLAC, or the audio track of a video. Both results come out as 44.1 kHz stereo WAV.
How long does it take?
The instant mode takes a second or two. Fast quality took about 4-6 seconds for a 3-minute song on a recent Mac in Chrome (plus the one-time download), and longer on older computers and phones. Best quality can take several minutes for a full song on a CPU, and is much faster in a browser with WebGPU. Before you start, the page shows an estimate for your device and song; once the first few seconds are processed it switches to the measured speed, with the time left. Keep the tab open until it finishes - the tab title shows the progress if you switch away.