A karaoke version of a song is the original recording with the singer taken out. You can make one from any MP3, WAV, FLAC or M4A file, and Mubert’s Karaoke Maker plays a 30-second preview before you sign up for anything. A free account gets the full backing track as MP3. This post is for anyone who has searched for a backing track at 11 p.m. before a party, an audition or a wedding toast, and it covers how to make the track, how to practise with it, where vocal removal fails and what you may do with the file afterwards.
Search results give you someone else’s karaoke version of the song
Type a song title plus “karaoke” into any search bar and you rarely get the record you know. The karaoke version you find is usually someone else’s guess at it. You get an upload in a different key. Or a two-minute intro somebody added. Or a man in a bedroom playing the whole thing on a keyboard set to “Strings 2”, with the lyrics in yellow Comic Sans.
Some of these are fine. Plenty are not, because they are re-recordings, and a re-recording is a different song wearing the original’s jacket. We have sung over a few. The worst one had the chorus in the right key and the verses a semitone lower (a sentence nobody should have to type).
YouTube’s own help pages treat that category separately. The page on what is eligible for Content ID has a list of ineligible references. It names “karaoke recordings, remasters, sound-alike recordings, and some dubbed content”. Content ID is YouTube’s system that matches uploads against reference files from rights holders. In plain words, this audio is too close to something else to count as the original. Fair enough.
So the argument of this post is simple. The backing track you want is the record you already know, every drum fill and guitar squeak in place, minus one voice. Since about 2019 you can make that karaoke version yourself, often faster than scrolling past the keyboard man.
Vocal removers split the mix with AI, and the tech went open source in 2019
A vocal remover is software that separates a finished song into parts. Engineers call it source separation. The model hears one stereo file and guesses which sounds belong to the voice and which belong to everything else. Then it writes two files instead of one.
The old trick was cruder. You flipped one stereo channel and subtracted it from the other. Whatever sat in the dead centre of the mix vanished. That did take the voice out. It also took out the kick, the bass and the snare, which sit in the centre too, so the karaoke version that came out was a ghostly hiss of reverb and hi-hats. Nobody sang over that twice.
Machine learning changed that. Deezer’s research team released Spleeter as an open-source tool in 2019, with models that split a song into two, four or five parts, and suddenly anyone with a laptop and some patience could pull a voice out of a finished mix. Meta’s researchers followed with Demucs, which separates drums, bass, vocals and “other”. It also has a two-stem option: vocals against accompaniment. That is, functionally, a karaoke button for people who enjoy the command line. The Demucs repository now says it is no longer maintained. Still, the idea is everywhere, because browser tools wrap models like these so you never open a terminal.
Karaoke began with backing tracks, and the first machine dates to 1967
The first karaoke machine, Shigeichi Negishi’s Sparko Box, dates to 1967, and it grew out of backing tracks. A small tangent, then. Before the Sparko Box, NPR’s obituary says, “karaoke-like activities involved the use of backing tracks”. The same report credits nightclub musician Daisuke Inoue with building his own machine independently in 1971. Negishi never secured a patent for his. The word comes from Japanese kara, empty, and ōke, short for orchestra, Merriam-Webster says. Fifty-odd years later, the empty orchestra still needs someone to empty it.
Anyway.
How to make a karaoke version with the Vocal Remover
The process takes three steps, and the tool asks for nothing but the song. The Vocal Remover accepts MP3, WAV, FLAC or M4A files. Upload the song, pick “Remove vocals” or “Extract vocals”, listen, then download. That is the whole karaoke version workflow.
Both choices produce both tracks. The button only decides which one plays first, so you cannot pick the wrong one. For a backing track, choose “Remove vocals” and it plays first. To get an a cappella for a remix, choose “Extract vocals”.
Before downloading, mute each track in turn. Every time. This is the quality check. Listen to the instrumental alone for traces of the singer. Then listen to the vocal alone and check whether a guitar line went off with it. Ten seconds of each tells you more than the waveform does.
Without an account you hear a 30-second preview. A free account gets the full-length tracks, and the free plan includes 20 minutes of songs per month, downloaded as MP3. WAV downloads start at the Creator plan. Plans change, so check the live pricing page rather than trusting a blog post (including this one).
Feed it the best file you have. The tool page says WAV or FLAC separates more cleanly than small re-encoded MP3s or audio ripped from a video. Compression throws away fine detail. Sadly, that detail is exactly what the model uses to tell a voice from a synth. A karaoke version made from a rip of a rip sounds like one. If you only have a video, save its audio in one of the four formats first.
Keep the singer as a quiet guide while you learn the song
The Karaoke Maker runs the same separation with a different goal. It is built for singing along, so the vocal track stays available as a guide. Instead of muting the original singer, you turn them down low and sing over them.
This is the most useful feature in the whole process, and people skip it. We did. On the first try we muted the vocal completely. Then we learned, live, that nobody in the room knew how the second verse started. The first verse was fine. The chorus was fine. Then the band played four bars of something. Four of us looked at each other like passengers on a train stopped between stations. Nobody moved.
A guide vocal fixes that. Practise the karaoke version with the singer at a low level for the first few runs. You hear the phrasing that way: where they breathe, where they come in late, which word they stretch. Then turn them down further. On the night, mute them, or leave a whisper of the original if you sing a duet part and need your cue.
The download is audio only. The tool page says so directly: “You get audio tracks, not a lyric video.” So there are no lyrics on screen; print them or keep them on a second device. The page doesn’t mention a key change either. If the song sits too high, you need a separate pitch-shifting step in an audio editor. Shift it more than a couple of semitones and the band starts to sound like it inhaled helium.
Vocal removal leaves ghosts in reverb, choirs and saxophones
Vocal removal is good now, but it still guesses. The Vocal Remover page names the usual suspects: long reverb on the vocal, choirs, and instruments in the singer’s range, like a lead synth or a saxophone. Each can leave faint traces of the voice.
The reason is physical. A sax in the same register as a tenor shares a lot of frequencies with him. Reverb spreads the voice across the stereo field for seconds after each word. The model has to decide whether that wash belongs to the singer or the room. Sometimes it splits the difference. Your karaoke version then carries a faint second singer, like a neighbour practising the same song through the wall.
The second limit surprises people more. The tool sends all singing to the vocal track, lead and backing vocals together. If the crowd normally sings the “oohs” or the call-and-response in the chorus, those go too. Maybe you want that. Now you get to sing them. It might also leave a strange hole where the hook used to be, so listen through every chorus before the night, not during it.
We once fed a separator a phone recording of a concert, the only copy of one particular arrangement. The crowd came through louder than the band. The model decided a thousand people singing counted as “vocals” and took most of the atmosphere with the singer. Hard to argue. Use a studio recording when one exists.
Sometimes the problem is a guitar part you want to play yourself. The same tool set also has a Stem Splitter, a Drum Remover and a Bass Remover. We tested six stem splitters on one song to see what each lets you download free.
Streaming apps turn the singer down, but they don’t give you a file
If you only need to sing in the living room, your music app may already do half the job. Apple announced Apple Music Sing on December 6, 2022. It adds a slider that controls the vocal level on millions of songs in the Apple Music catalogue. It works on iPhone, iPad and the Apple TV 4K, with lyrics that move in time with the voice.
Apple’s own footnote is honest about the limit: “The vocal slider adjusts vocal volume, but does not fully remove vocals.” Apple also excluded Voice Plan subscribers. And because it is a playback feature, it gives you nothing to take anywhere else. It is a karaoke version on a leash.
That matters more than it sounds. An audition, an open mic or a wedding DJ wants a file, and so does a choir director, a vocal coach, or the friend running the speaker at a barbecue who does not share your subscription and never will. A streaming slider is a remote control. A karaoke version is a file you keep. It sits on a phone, a USB stick or in an email that arrives before the event.
Lyric videos on video sites have the same problem. They are fine for practice in the kitchen. They are a bad plan at a venue with patchy wifi and a bartender who wants the aux cable back.
So use the slider for practice, if you have it, and make a file for the performance. Name the files properly, too. Two downloads with nearly the same name look identical on a phone at a party. You will play the one with the singer in it, in front of everyone. Ask us how we know.
Singing over a karaoke version at home is not the same as uploading it
Removing the vocals does not change who owns the music. The backing track is still the label’s recording with one part turned off. The song underneath (the melody, the chords, the lyrics) still belongs to its writers and publishers. A vocal remover changes the mix. It does not change the paperwork.
For a party, a rehearsal or a gig where the venue handles licensing, that is usually where the question ends. Uploading is different. A video of you singing over a homemade karaoke version still contains the original recording. Expect a rights holder’s matching system to spot it. Quickly.
Streaming yourself singing counts as publishing, just faster. A live stream carries the same original recording as an upload, so the platform’s music rules apply to it too.
Wedding films are the classic trap. Everybody sang along at the reception, and yet the first dance song is the one track the video can’t use without permission.
Covers with your own band raise a separate question, because the composition is still someone else’s. If the upload matters, for a channel you earn from or a client video, check the platform’s rules for covers. When in doubt, ask whoever owns the song. We are a music-tech crew, not your lawyer, and copyright rules differ by country.
One in four checked tracks has a claim, and a karaoke version can’t be a Content ID reference
About one in four tracks people run through our copyright checker come back with a rights holder’s claim. The share is 24.3%, down slightly from 27% (Mubert’s own data, last 28 days). People run that check because they suspect a problem, so it does not describe all music. It does show how often the suspicion is right.
The money doesn’t flow your way either. YouTube will not accept a karaoke version as a Content ID reference, per the eligibility page quoted earlier. So you cannot register your instrumental and collect each time someone else sings over it. We checked twice, out of hope.
A claim on a video you care about is not always the end. In some regions YouTube can now swap the claimed song for an AI-written one, which beats deleting the whole video.
Some people searching for a karaoke version actually want something else: a track to sing their own words over, like a birthday song for a colleague or a jingle for the shop down the street. For those, an original instrumental in the right mood skips the borrowed-record question, and a generated track written to the exact length is a cleaner starting point. The licence scope depends on the plan, so read it before you rely on it.
A checklist for your next karaoke version
- Find the best-quality copy of the song you have, ideally WAV or FLAC.
- Upload it to the Vocal Remover or Karaoke Maker and play the free 30-second preview first.
- Mute each track once to check for vocal ghosts and missing instruments.
- Sign in with a free account and download the full karaoke version and the vocal track.
- Practise three runs with the singer turned down low, then one run with them muted.
- Listen through every chorus for missing backing vocals before the night.
- Print the lyrics, because the tool gives you audio, not a lyric video.
- Check the platform’s rules before uploading anything with the original recording in it.
The keyboard man in the bedroom has done nothing wrong. He played every note himself, Comic Sans and all. He just doesn’t have to be your band any more.
Who we are
