How to remove vocals from a song
- Add your song. Drop an MP3, WAV, M4A, FLAC or OGG file into the box above, or tap it to choose a file from your device.
- Choose the output. Pick Instrumental for a karaoke track, Acapella for the isolated voice, or Both.
- Remove the vocals. The first time, your browser downloads the AI model once (about 180 MB). After that it starts instantly, even offline.
- Download. Preview the instrumental and the vocals, then save them as MP3 or WAV.
Powered by HTDemucs v4 AI
The vocal remover uses HTDemucs v4 (Hybrid Transformer Demucs), the open-source music separation model from Meta AI Research. It's the same family of models behind many professional stem-separation apps. It was trained on thousands of multitrack recordings, so it has learned what a voice sounds like, not just where it sits in the stereo mix. That means it can:
- remove lead and backing vocals, including ones panned to the sides or drenched in reverb,
- keep centre-panned instruments like bass, kick, snare and lead synths intact in the instrumental,
- work on mono recordings, live recordings and old records that the traditional "phase cancellation" trick can't handle.
The model combines two views of the audio. A waveform branch captures sharp transients and timing, while a spectrogram branch captures pitch and harmonics. A cross-domain transformer lets each branch learn from the other across several seconds of context. The song is processed in overlapping 7.8-second windows that are blended together seamlessly.
Runs on your device, not our servers
Unlike most online vocal removers, your song is never uploaded. The AI model runs inside your browser using WebGPU on your graphics card when available (Chrome, Edge, and recent Safari and Firefox), or on the CPU otherwise. On a recent graphics card, a 4-minute song usually takes 1–3 minutes. Older laptops with integrated graphics can take 5–10 minutes, and CPU-only processing takes longer. The progress bar shows the time remaining.
Other engines
If you can't download the model right now, choose Instant (a fast spectral method, no download) or Classic phase cancellation (the old karaoke-machine trick). They're quicker, but their quality is noticeably lower than the AI engine.
Tips for the best results
- Start with good quality. WAV, FLAC or 256–320 kbps MP3 files give the cleanest separation. Heavily compressed files contain artefacts the AI can't undo.
- Use a modern browser. The latest Chrome or Edge on a computer with a graphics card gives the fastest results.
- Save as WAV if you plan to mix or edit the results further.
- Need more than vocals? The Stem Splitter uses the same AI to separate drums, bass and other instruments too.
- Change the key. Put the instrumental through the Audio Scale Changer to sing in a key that suits your voice.
Frequently asked questions
Is this AI vocal remover free?
Yes. It's completely free with no sign-up, watermark, or limit on how many songs you process.
Is my song uploaded to a server?
No. The AI model is downloaded to your browser and runs on your own device, so your music never leaves your computer or phone.
Why is there a download the first time?
The AI model's weights (about 180 MB) must be on your device to run locally. They're saved in your browser's storage, so later uses start instantly without downloading again.
How long does it take?
With a recent graphics card (WebGPU), usually 1–3 minutes for a 4-minute song. Older integrated graphics can take 5–10 minutes, and CPU-only processing takes longer. The progress bar shows an estimate while it works.
Which AI model is used?
HTDemucs v4, the hybrid transformer version of Demucs from Meta AI Research, released under the MIT license.
Does it work on phones?
It works on recent phones with enough memory, but it's much faster on a laptop or desktop. On phones, the Instant engine is a quick alternative.