How to Make a Karaoke Track from Any Song: Three Ways, Free in the Browser
Take a record, pull the singer out, keep the band: that is a karaoke track, and the word itself says so. Karaoke is Japanese, kara for empty and oke clipped from okesutora, orchestra, so an empty orchestra, a band with nobody at the microphone. Daisuke Inoue built the first machine for it in Kobe in 1971, a coin-operated box that played his own backing tapes for a regular at the bar where he drummed, and he never patented it. Fifty years on, the tapes are gone and the question is how to get the singer out of a song you already have on your phone. There are three ways. One of them needs luck, one takes ten seconds and sounds like it, and one takes a few minutes and works on nearly anything.
Which word means what
People use karaoke track, instrumental and backing track as if they were one thing, and for the audio they often are. The difference is what else is missing and what you do with the file.
| Term | What is removed | What stays | Used for |
|---|---|---|---|
| Karaoke track | The lead vocal, and in practice the harmonies with it | Drums, bass, guitars, keys | Singing over it |
| Instrumental | Lead vocal and harmonies | The same band | A release, a cover video, busking |
| Backing track | The vocal plus whatever the band plays live | What the band cannot cover | A live set |
| A cappella | Everything except the voices | Lead, doubles, harmonies | Remixes, vocal study |
| Lyric video | Nothing from the audio | Sound plus words on a screen | The screens at a karaoke night |
The lyric video is a separate job. The sound comes first, and the words and their timing get added in a video tool afterwards; nothing on this page puts text on a screen.
Way 1: somebody already made one
Record labels have been releasing instrumentals as B-sides since the days of the 7-inch, and the big karaoke catalogues, Karaoke Version is the one most singers know, sell re-recorded versions of the hits with the lead vocal left out. If the song is a chart single from the last forty years, there is a fair chance one exists, and it will be clean, because nothing was removed: the voice was never printed. Two catches. An album cut, a live recording or anything outside the charts usually has no official version at all, and a karaoke catalogue version is a cover, played by a session band in a different studio, so it is not the record you know, however close it gets.
Way 2: the Audacity trick
Audacity has carried an effect called Vocal Reduction and Isolation since version 2.1.2, in 2016, written by Robert Hänggi as a Nyquist plug-in. Open the Effect menu, pick it, choose Remove Vocals in the Action box, leave the Strength at 1 and the two cut-off frequencies at 120 Hz and 9000 Hz, press Apply, and ten seconds later the voice is quieter. What it does is subtract one channel from the other, which cancels whatever was mixed dead centre at equal level in both channels; on most records that is the lead vocal.
It is also the kick drum, the snare, the bass, and the guitar solo on a great many rock records, because engineers put all of those in the centre too, and the effect cannot tell a singer from a snare. On a sparse acoustic song you can get away with it. On a dense mix the low end goes thin, the drums lose their punch, a halo of the voice stays behind in the reverb, and the whole thing folds toward mono. It costs nothing and it is worth trying once, so you know what it sounds like.
Way 3: stem separation
The third way did not exist in a usable form until 2019, when Deezer published Spleeter and Facebook's research group published Demucs, both of them neural networks trained on multitrack recordings to listen to a mix and write out its parts. Demucs went through a hybrid version in 2021 and a transformer version in 2022, and the six-source variant of that model, the one this site runs, returns vocals, drums, bass, guitar, piano and a sixth stem for whatever is left. It does not care where the voice sits in the stereo field and it cancels nothing; it estimates each part on its own, the way you would if someone asked you to hum the bass line.
Making the karaoke track with the karaoke maker goes like this:
- Drop in the song, MP3, WAV, FLAC or M4A, up to 10 minutes. A lossless file splits cleaner than a 128 kbps MP3, because the model has more to listen to.
- Wait. Two to five minutes on a cloud GPU; the queue position is on screen.
- Pull the vocal fader to the bottom. The other five stems keep playing.
- Rebalance if you like: a piano that carries the melody you sing against can come up a couple of dB, a loud rhythm guitar can come down.
- Sing along in the browser, or export a 24-bit WAV at the sample rate of your own file.
Every voice on the record lands in that one vocal stem, lead, doubles and harmonies together, so a stacked chorus disappears as a stack. No model separates the lead from the harmonies from a stereo mix; that needs the multitrack, and the multitrack is in a vault somewhere.
Keep a little of the singer
The vocal stem is a fader, not a switch, and that is the thing a bought karaoke version cannot give you and the Audacity trick has thrown away. At a quarter of its level the singer is still there under the band, marking every entry while you learn the song; at zero it is karaoke. Switching between the two is one move, and nobody has to re-upload anything.
What stays in, whichever way you go
Reverb tails, first of all. The reverb printed on the voice carries an echo of it, and that echo is the room rather than the singer, sitting twenty decibels or more below the band. Then the places where the voice overlaps a piano or a clean guitar in pitch, where a trace of each stays in the other. And live recordings, which hand the model audience noise, stage bleed and a wet room all at once, come out rougher than a studio mix with a dry vocal in the middle. None of it is audible once you sing.
The key, and the one
A karaoke track from a record is in the record's key, which is the singer's key, not yours. If the chorus sits too high, move the file, not the song: CountIn, the sister tool to this one, shifts the key by up to 7 semitones with the tempo untouched, and puts a count-in of 1 to 8 bars in front, so the first line still has a cue after the band stops playing it for you. Both come out in one export at countin.band.
Frequently asked questions
Is a karaoke track the same as an instrumental?
The audio can be identical, the song minus the lead vocal. The word follows the use: an instrumental is for a release or a cover, a karaoke track is for singing over, which is why a guide vocal matters for one and not the other.
Can I make a karaoke video with the lyrics on screen here?
No. This site makes the sound. The WAV goes into a video editor or a karaoke app, and the words and their timing are added there.
Does it work on any song?
On any file up to 10 minutes, yes. How clean it comes out depends on the mix: a modern studio recording with a centred vocal is the easy case, a live album or a voice buried in distorted guitars leaves more of the band in the stem.
Is it free?
Two songs a day per person, no account, no watermark, a full-length 24-bit WAV export, while the site is in beta. There is a site-wide daily cap on top, because separation runs on paid GPUs.
What happens to the backing vocals?
They go with the lead. All the voices on the record land in the vocal stem, so muting it removes the harmonies as well.