How Does a Vocal Remover Work? Inside Our Browser Engine
A vocal remover works by sorting a song’s sound into the parts that belong to the voice and the parts that belong to the music, then rebuilding them as two separate tracks. Our tools do this inside your browser in one of two ways: with HTDemucs, an open-source AI model from Meta, or with a Quick mode that pulls out whatever sits in the center of the stereo mix. Your song never leaves your device.
VocalRemover is a free set of tools built and maintained by the team behind this site. We wrote this page from the tools’ own code and from tests we ran in Chrome in November 2026. Last updated: November 2026.
How to Use It in Four Steps
Every tool on the site follows the same four steps, and nothing gets installed.
1. Choose your tool
The Vocal Remover gives you both the vocals and the music. The Karaoke Maker, Acapella Extractor, Instrumental Maker and Background Music Remover each open with the settings and tracks that suit their job.
2. Add your file
Drop a song or video onto the tool, or browse for it. It accepts MP3, WAV, FLAC, M4A, OGG, OPUS, MP4, MOV and WebM files up to 150 MB and 8 minutes long. Stereo files give the best results.
3. Pick an engine
AI separation gives the cleanest result and handles files up to 6 minutes. Quick finishes in seconds and handles the full 8 minutes. The tool starts on AI when your browser can use your graphics card.
4. Compare and download
Play the new tracks against your original, adjust the options if you need to, and download the track you want as a WAV file. There’s no sign-up and no watermark.
Two Ways a Vocal Remover Can Work
Vocal removers use one of two methods, and our tools offer both. Center extraction relies on lead vocals usually being mixed in the center, while AI source separation uses a model trained to recognize voices.
The classic form of center extraction is phase cancellation. It flips one channel and adds it to the other, so anything identical in both channels cancels out, and that usually includes the vocal. The catch is that the bass, kick drum and snare often sit in the center too, so they disappear with the voice and the result loses punch.
Our Quick mode applies the same idea band by band, as described below. AI mode doesn’t depend on where the voice sits in the mix at all.
| Feature | Center extraction (our Quick mode) | AI source separation (our AI mode) |
|---|---|---|
| How it finds the voice | Looks for sound that the left and right channels share | A trained neural network picks out the voice by its sound |
| Speed | Seconds | Minutes or much longer, depending on your graphics card |
| Works best on | Studio songs with the voice in the center | Songs that Quick mode leaves messy |
| Common weakness | Centered bass, kick drum and snare can get caught too | Heavy reverb and dense mixes can leave artifacts or faint traces |
| File limit | 8 minutes | 6 minutes |
Inside the AI Engine
Our AI mode runs HTDemucs, short for Hybrid Transformer Demucs, a model released by Meta’s research team under the open-source MIT license. It analyzes the sound both as a waveform and as a spectrogram, which is a picture of which frequencies play at each moment, and uses Transformer layers to tell the sources apart. HTDemucs splits a song into four parts, called stems: vocals, drums, bass and everything else. Meta reports a score of 9.00 dB SDR, a standard measure of separation quality, on the MUSDB HQ test set (Meta, Demucs on GitHub).
Here is what happens after you drop in a file:
- Decode. Your browser turns the file into raw left and right audio at 44,100 samples per second.
- Load the model. The first time you use AI mode, your browser downloads the model file, about 172 MB, from Hugging Face. It stays in your browser’s cache, so later runs skip the download.
- Separate. The model predicts the four stems. It runs on your graphics card through WebGPU when your browser supports it, and on your processor with up to four threads when it doesn’t.
- Rebuild two tracks. The tool keeps the vocal stem as your vocal track and adds the drums, bass and other stems back together as your instrumental.
- Save as WAV. Both tracks are saved as 16-bit stereo WAV files, with a peak limiter so they don’t clip.
How Quick Mode Works
Quick mode is a frequency-domain center-extraction engine built for speed and for older devices. It cuts the song into short frames of 4,096 samples, which gives about 2,000 frequency bands per frame. In each band, it measures how much sound the left and right channels share and treats that shared part as the voice.
It only searches for the voice between about 70 Hz and 11,500 Hz, so the deepest bass and the highest cymbal shimmer stay in the instrumental. The Strength presets set how hard it pulls out the center.
What We Measured
Quick mode separated our 60-second test file in about 20 seconds, while AI mode was far too slow to finish on our older laptop. We ran both engines in Chrome on a Windows laptop with an integrated Intel Gen-9 graphics chip, using WebGPU. The test file was a 60-second stereo WAV (10.1 MB) that we generated ourselves, with a voice-like 220 Hz tone in the center, a chord panned mostly left and noise-based drum hits panned mostly right.
Our test voice sits perfectly in the center, which favors Quick mode, so expect less clean results from real songs. AI mode was very slow on this older integrated chip, and the page sometimes stopped responding for up to 45 seconds. The site says AI speed depends on your graphics card, but we haven’t timed a newer one, so we won’t promise a figure.
| Test | Result |
|---|---|
| Quick mode time | About 20 seconds for the 60-second file |
| Quick mode instrumental | Matched the original music at 0.93 correlation, up from 0.63 for the full mix (0.14 with the voice) |
| Quick mode vocal track | Matched the original voice at 0.95 correlation, up from 0.78 for the full mix (0.12 with the music) |
| Output files | 44,100 Hz WAV, 60.0 seconds each: stereo instrumental (10.1 MB) and mono vocals (5.0 MB) |
| AI mode on this laptop | With the model already cached, separation started about 4 minutes after we dropped the file. At 8 minutes (45%), the tool estimated about 40 minutes left, so we cancelled. |
Choose the Right Tool
All five tools use the same two engines, but each one opens with the tracks and options that fit its job.
| Tool | What you get | Extra option |
|---|---|---|
| Vocal Remover | Vocals and instrumental | Remove vocals or remove music, plus keep bass and kick |
| Karaoke Maker | A backing track without the lead singer | Guide vocal |
| Acapella Extractor | The vocal on its own, plus an optional instrumental | Cut low rumble |
| Instrumental Maker | The music without the voice | Keep bass and kick |
| Background Music Remover | Clear speech from videos and podcasts | Speech focus |
What the Controls Do
These options fine-tune Quick mode, and most of them work by changing the frequency range it searches for the voice.
- Strength: Three levels set how much sound counts as the voice. They’re called Light, Balanced and Strong in most tools, Keep, Balanced and Remove more in the Karaoke Maker, and Gentle, Balanced and Tight in the Acapella Extractor. Stronger settings remove more of the center, which can take more of the music with it.
- Keep bass and kick: This raises the bottom of the search range from 70 Hz to 150 Hz, so centered bass and kick drums stay in the instrumental.
- Guide vocal: In the Karaoke Maker, this leaves a faint trace of the singer in the backing track, so you can follow the melody.
- Cut low rumble: In the Acapella Extractor, this raises the low cutoff, so less bass and kick leaks into the vocal.
- Speech focus: In the Background Music Remover, this lowers the top of the search range from 11,500 Hz to 7,500 Hz, which keeps the search on the range where speech sits.
- Compare at the same moment: Play one track, then tap another. Playback jumps to the same point, which makes small differences easy to hear.
Why Some Songs Separate Better Than Others
Studio songs with a dry, centered lead vocal separate best. Heavy reverb on the voice and dense mixes where many instruments overlap it can leave artifacts or faint ghosting, even in AI mode.
Quick mode has two extra blind spots. It only finds what sits in the center, so vocals panned to the sides stay in the instrumental. Mono files are a problem too: every sound is identical in both channels, so Quick mode can’t tell the voice from the music. For songs like these, use the “Not clean enough? Use AI separation” link on the results screen.
Your Files Stay on Your Device
Your audio is never uploaded. Your browser opens the file, the separation runs inside the tab, and the downloads are built on your device. AI mode does make two outside requests: it fetches the model from Hugging Face on the first run, and it loads the ONNX Runtime Web and demucs-web code from jsDelivr. Neither request includes your audio.
How It Works FAQs
How long does vocal removal take?
Quick mode takes seconds: our 60-second test file finished in about 20 seconds. AI mode takes much longer and depends on your graphics card. On our laptop with older integrated Intel graphics, separation started about 4 minutes in, and at 8 minutes the tool still estimated about 40 minutes left. The first AI run also downloads the 172 MB model once.
Does the vocal remover use AI?
Yes. AI mode runs HTDemucs, an open-source AI model from Meta, directly on your device, and the tool picks it by default when your browser can use your graphics card. Quick mode is not AI. It uses signal processing to pull the shared center out of the stereo mix, which makes it fast and suitable for older devices.
Why are the downloads in WAV format?
WAV keeps the full quality of the separated audio with no extra compression, and audio editors widely support it. Our tools save 44,100 Hz WAV files. AI mode tracks are 16-bit stereo, while Quick mode’s vocal track is mono. If you need a smaller file, you can convert the WAV to MP3 afterward.
Which browsers does it work in?
The tools work in current versions of Chrome, Edge, Firefox and Safari, on desktop and mobile. AI mode runs fastest in browsers that support WebGPU, which lets the model use your graphics card. Without WebGPU, it falls back to your processor and runs very slowly, so the tool recommends Quick mode instead.
Do I need to keep the page open?
Yes. The separation runs inside your browser tab, so keep it open until your tracks appear. Closing or reloading the page clears any results you haven’t downloaded. Nothing is stored on a server, so download the tracks you want before you leave.
Can I publish a song after removing the vocals?
Usually only with permission from the rights holder. Removing the vocals doesn’t change who owns the song. The instrumental and the vocal track still come from the original artist’s recording, so publishing them, for example in a video or a cover release, usually needs that permission.
Ready to Try It?
Drop in a song, and Quick mode will give you separated tracks in seconds. If the result isn’t clean enough, switch to AI separation on a device with a capable graphics card. It’s free, and your file never leaves your device.