Paste a song into a vocal remover, wait a minute, and the singer is gone. It feels like a trick, and for a long time it was one: the old methods either cancelled the centre of the stereo image or cut a band of frequencies, and both wrecked the rest of the music. Modern tools do something different. This guide explains what, without equations.
A song is a picture
Audio is a wave: air pressure going up and down thousands of times per second. That is awkward for a computer to reason about, so the first step in almost every separation system is to turn the sound into a picture called a spectrogram. Time runs left to right, pitch runs bottom to top, and brightness shows how much energy there is at each pitch at each moment.
In that picture a voice looks like a set of wavy horizontal lines: a fundamental pitch and its overtones, sliding as the singer moves between notes. A kick drum is a short vertical smear at the bottom. A hi-hat is a fine mist at the top. A piano chord is a stack of straight lines that start together and fade. The instruments overlap, but they have shapes, and shapes are something a neural network can learn to recognise.
The model paints a mask
Given the spectrogram of a mixed song, the network's job is to output a second picture the same size, called a mask, where each pixel says how much of the energy at that time and pitch belongs to the voice. Multiply the original picture by the mask and you have the voice's spectrogram; multiply by the inverse and you have everything else. Turn those pictures back into waves and you have two audio files.
A stem splitter does the same thing several times over, producing one mask each for vocals, drums, bass, piano and "other".
Newer models add a second view: alongside the spectrogram they also look at the raw wave directly, which helps with sharp transients such as drum hits that get blurred in a spectrogram. Some of the engines you see in tools, such as MDX-Net and Demucs, differ mainly in how they combine these two views and how large they are.
How it learned what a voice looks like
Nobody programmed rules like "a voice has overtones spaced like this". The network was trained on thousands of songs where the separate tracks were available: the mix on one side, the isolated vocals, drums, bass and other stems on the other. For each song it guessed a mask, compared its guess with the real isolated stems, and adjusted millions of internal numbers slightly to do better next time. Repeat for weeks on powerful hardware and the result is a model that has, in effect, heard more multitrack sessions than any human engineer.
That training set is also the source of its limits. The model is best on the kinds of music it heard most, which is modern pop, rock, hip-hop and electronic music with clear studio production.
Why some songs separate cleanly and others do not
Reverb and delay. The effects on a voice are made of the voice, smeared in time and mixed with the room. Part of that smear looks like the instruments underneath it, so the model has to guess where the voice ends and the room begins. Dry vocals separate cleanly; cathedral-sized reverbs leave a ghost.
Doubled and layered vocals. Harmonies, doubles and stacked choruses are all "voice" to a two-stem model, so they all leave with the lead. If you want to keep the harmonies, you need a model trained specifically to remove only the lead, which is what a "karaoke" model is.
Voices used as instruments. A vocoder, a talk-box or a heavily pitch-shifted vocal chop shares more shape with a synth than with a singer. Some of it will stay in the instrumental.
Overlapping registers. A distorted rhythm guitar and a female belt live in the same region of the picture. Where they coincide, the mask is a compromise.
Lo-fi sources. A 96 kbps stream or an old cassette transfer has already lost the fine detail the model uses to tell sources apart.
What "engine" and "quality" settings really change
Vocal removers usually offer a standard engine and one or two higher-quality ones. The difference is not a filter; it is a different, larger model, sometimes several models whose outputs are averaged. A larger model makes better guesses at the ambiguous pixels, which shows up as less vocal ghost in the instrumental and fewer instrument fragments in the acapella. The cost is time: an ensemble of two big models can take several times longer than the standard one.
On Mazmazika the standard engine gives up to five stems quickly; the High Quality models (Best Vocals, Best Instrumental, Karaoke) and the Ultra engine, which runs two models and merges the best of each, are the larger options, and the tool page lists which of them your plan includes. The output bit rate rises from 128 to 320 kbps on those engines as well, which matters if the file will be played on a large system.
Why two stems are cleaner than five
Every extra stem is another boundary the model has to draw. A "voice or not voice" model draws one boundary, and the instrumental keeps the original mix's cohesion. A five-way model draws four boundaries, and small errors at each one add up when you play the stems together again. That is why the karaoke guide recommends two stems even though five are available.
What the model cannot do
It cannot invent what was never recorded. If the singer's breath completely covered a quiet piano note, that note is gone in the instrumental. It cannot separate two singers of the same range singing in unison into two people. And it cannot remove copyright: the separated stems are still the original recording, legally speaking, so the rules about publishing and performing still apply.
Where the technology is going
Research models keep getting larger and are now trained with more stems (guitar and piano as separate targets, for instance) and on more genres. The gap between a free browser tool and a studio workstation has mostly closed for two-stem separation; the remaining differences are in edge cases, speed, and how the tool lets you work with the result afterwards.
Try it with your own ears
The fastest way to understand all of this is to run a song you know well through the vocal remover, mute the vocal stem, and listen to the chorus with headphones. You will hear exactly the effects described here: a clean verse, a faint ghost in the reverb tail, and a chorus where the harmonies left with the lead. Then read vocal remover vs stem splitter to decide which mode fits your next project.