At a glance
- Cost
- 100% free, no account
- Model
- Meta HTDemucs v4 neural network (instant and classic modes too)
- Privacy
- Songs are processed locally and never leave your device
- Accepts
- MP3, WAV, FLAC, M4A, OGG or a video's soundtrack
- You get
- Karaoke instrumental, isolated vocals, or both (MP3/WAV)
- Setup
- AI model downloads once (~180 MB) and is cached
Step by step
- Pick a song. Any common format works: MP3, WAV, FLAC, M4A, OGG. A higher-quality source gives a cleaner split.
- Decide what you need. Keep Both to receive the instrumental and the acapella, or choose just one.
- Let the AI work. The very first run fetches the model (roughly 180 MB) into your browser's storage. Every run after that starts straight away.
- Listen, then save. Compare the two tracks in the built-in players and download them as MP3 or WAV.
What's happening under the hood
Older vocal removers relied on a trick: most singers are mixed in the middle of the stereo image, so subtracting the right channel from the left cancels them out. Unfortunately it also cancels the bass, kick and snare, and it fails completely on mono or reverb-heavy recordings.
Full Audio Tools takes a different approach. It runs HTDemucs v4, a hybrid transformer model published by Meta AI Research and trained on a large collection of multitrack studio recordings. Instead of guessing from stereo position, the network has learned what singing actually sounds like: its harmonics, its consonants, the way it moves. It predicts the vocal part directly and leaves everything else untouched.
- Two views of the music: one branch reads the raw waveform (great for timing and attack), another reads the spectrogram (great for pitch and tone).
- A transformer in the middle lets both branches share context across several seconds of audio.
- Seamless stitching: long songs are processed in overlapping windows that are cross-blended, so you won't hear joins.
Speed and hardware
The model runs on your graphics card through WebGPU whenever your browser supports it (current Chrome, Edge, Safari and many Firefox builds). A dedicated or recent integrated GPU typically finishes a 4-minute track in 1–3 minutes. Older laptops take longer, and without WebGPU the CPU does the work, which is noticeably slower. A live estimate of the remaining time is shown while it runs.
In a hurry? The Instant engine gives a rough result in seconds without any download, and Classic phase cancellation is available for comparison.
Getting the cleanest separation
- Start from the best copy you have. Lossless or 320 kbps files keep the detail the AI relies on.
- If you'll mix the result in a DAW, export as WAV so no extra compression is added.
- For separate drums and bass too, use the Stem Splitter.
- Want to sing it lower or higher? Run the instrumental through the Key Changer.
Common questions
Does the vocal remover cost anything?
No. It's free to use as often as you like, with no account and no watermark on the results.
Is my music sent anywhere?
No. Only the AI model is downloaded to your browser. Your song is processed entirely on your own device and never uploaded.
Why does the first run take longer?
The first time, your browser has to download the neural network (about 180 MB). It's then stored locally, so later runs skip that step.
Will it remove backing vocals too?
Usually, yes. The AI recognises voices wherever they sit in the mix, so harmonies and doubled vocals generally end up in the acapella track as well.
Can I use the results publicly?
Only if you have the rights. Separating a song doesn't change its copyright, so get permission before publishing covers or remixes of someone else's music.