How to remove vocals from a song without ruining the music

01

How to remove vocals from a song in four steps

The whole job takes a few minutes, most of it waiting for the first model download. You need a computer or phone with a current browser and the song as a file, for example an MP3 from your library or a video you recorded.

1. Choose the song

   Drop the file into the box or press “Choose a song”. MP3, WAV, M4A, AAC, FLAC, OGG, Opus, AIFF and the audio track of most videos work, up to 10 minutes and 300 MB.

2. Press “Remove vocals”

   The first time, the page downloads the AI model once (171 MB) and keeps it in your browser. Then it works through the song in parts of about eight seconds, and the progress line shows how many parts are done and roughly how long is left.

3. Listen to the result

   When “Your stems are ready” appears, the “Instrumental” preset is already selected. Play it, jump through the song and listen to quiet passages, where leftovers are easiest to hear.

4. Download the instrumental

   Press “Download instrumental”. You get a stereo WAV file at 44.1 kHz and 16-bit, the same length as the original song, named after the song with “instrumental” added.

02

Why the old karaoke trick made songs sound hollow

Before AI separation, the usual way to remove vocals was phase cancellation. Lead vocals are usually mixed to the middle, so they are identical in the left and right channel. Flip one channel upside down, add both together, and everything in the middle cancels out.

The trouble is that the voice is not the only thing in the middle. Kick drum, snare and bass often sit there too, so they vanish or shrink along with the singer. Reverb on the voice is spread across the stereo field, so it survives as a ghostly echo. And the result is effectively mono. That is why old karaoke versions made this way sound thin.

An AI separator works differently. The Vocal Remover uses Hybrid Transformer Demucs (htdemucs), an open-source model from Meta’s research team that was trained on many songs with known, separate tracks. It listens to the whole mix and rebuilds four tracks, called stems: vocals, drums, bass and other instruments. The instrumental is simply drums, bass and other added together, so the bass and drums keep their weight and the stereo picture stays intact.

03

Which file gives the cleanest instrumental?

The model can only work with what is in the file. Compression that throws away detail, background noise and room sound all make the separation harder. As a rule of thumb:

04

Why can I still hear the voice? Leftovers and what to do

Any separator, AI or not, has limits. These are the situations where you are most likely to hear a trace of the singer, and what helps:

• Reverb and echo on the voice: the dry voice is removed, but part of its reverb tail can remain as a soft wash. It is usually covered when the song plays at normal volume.
• Backing choirs and stacked harmonies: when they are mixed like a pad, the model may file part of them under other instruments. Switch on only “Other instruments” to check.
• Instruments that sing: a lead synth, violin or saxophone in the vocal range can partly end up in the vocals stem, so the instrumental loses a little of it.
• Shouted ad-libs and spoken intros: these are voices too, and they are usually removed with the rest.
• If a trace bothers you, use the switches: keep drums, bass and other at 100%, and decide whether you prefer a little leftover voice or a slightly thinner instrument.

05

Remove vocals from an MP3, a video or a voice memo

The Vocal Remover accepts whatever your browser can decode, so you rarely need to convert anything first. An MP3 from a music library works directly, and so do MP4, MOV, WebM and MKV videos: the page reads only their audio track and gives you WAV files, not a new video.

If you want to use the instrumental under a video again, put the WAV file into your video editor and line it up at the start of the clip. Because the file is exactly as long as the original audio, it stays in sync when both start at the same point.

If the browser cannot read the audio in a file, the page says so. In that case, export an MP3, WAV or M4A version from the program that made the file and try again. Songs longer than 10 minutes need to be cut first.

06

How long it takes, and where your song goes

Speed depends on your device. In Chrome with WebGPU, the browser runs the model on the graphics chip: in our measurement on a desktop PC, a 27-second clip took 6.6 seconds, and a three-minute song finishes in well under a minute. Without WebGPU, the model runs on the processor and a song takes several minutes. The page tells you when that is the case.

Your song is never uploaded. The browser decodes it, the model separates it on your device and the downloads are created there too. Only the model and the ONNX Runtime engine that runs it are downloaded, and after the first time the model comes from your browser’s storage, so the tool also works offline.

07

What you may do with the instrumental

Removing the vocals does not change who owns the song. Using the instrumental for your own karaoke night, singing practice or study is fine. Publishing it is different: a remix, a cover backing track on a streaming platform or a video with the music needs the rights to the recording and the composition, because both still belong to their owners.

The U.S. Copyright Office explains that preparing derivative works is one of the owner’s exclusive rights, and that fair use is decided case by case. This guide is not legal advice; when in doubt, ask the rights holder or a lawyer before you publish.