On this page
Every vocal remover produces something slightly wrong. A trace of the singer in the chorus, a hollow quality in the drums, a word or two surviving in the instrumental. People assume they picked the wrong app.
They did not. The limitation is in the problem, not the product, and understanding it tells you which songs are worth attempting.
Separation is a guess, not an undo
A finished song is a single stereo file. Two channels, and in each one, a single stream of numbers describing how a speaker cone should move.
That file does not contain a vocal track. It contains the sum of everything, already mixed down. The voice, the drums, the bass, the guitars and the reverb were added together into one waveform, and addition loses information: given only the total, you cannot recover which numbers produced it.
So a separator is not undoing the mix. It is listening to the result and estimating what must have been in it — using a model trained on a great many songs where it was shown both the mix and the true separate parts, and learned what a voice tends to look like inside one.
That is a genuinely impressive thing to be able to do. It is also, by construction, a guess. And the quality of a guess depends entirely on how distinguishable the voice was in the first place.
The old trick, and why it was worse
Before machine learning, the common approach was phase cancellation: lead vocals are usually mixed centre, meaning identical in the left and right channels, so subtracting one channel from the other removes anything that is identical and leaves what differs.
It works, and it is terrible, because centre is a crowded place. Kick drum, bass and snare are usually centre too, so they vanish with the vocal. Meanwhile any reverb on the voice is deliberately different in each channel to create width, so it survives. The classic result is a thin, bass-less track with a ghost of the singer's reverb still floating over it.
Modern separation identifies sounds by what they sound like rather than where they sit, which is why it keeps the bass. The remaining errors have a different cause.
What makes a song hard
The model has to decide, for every frequency at every instant, how much of that energy belongs to the voice. It is accurate where the voice is distinguishable and guesses where it is not.
Overlap. A voice and an instrument occupying the same frequency range at the same moment is an ambiguous case. Some of the energy is assigned wrongly, which is heard as bleed in one stem and a hole in the other.
Masking. When something much louder covers the voice, the voice's detail is barely present in the mix. The model has little to work from, and a reconstruction from almost nothing sounds like a reconstruction.
Reverb and effects. Reverb spreads a voice across time and across the stereo field, so the tail is no longer localised with the voice that produced it. Long reverb tails are among the most common things left behind in an instrumental.
Layering. Doubled vocals, harmonies, choirs and vocal chops all mean several voices at once with different characteristics, and the chorus — the part you most want clean — is usually the most layered part of the song.
Live recordings. Everything bleeds into every microphone before the mix even happens.
A dry, sparse, modern pop production with one lead voice is close to the easy case. A dense live rock recording with layered harmonies is close to the hard one. No app changes that ordering; they differ in how well they handle the middle.
Two stems are cleaner than four
If you only need vocals and instrumental, ask for two stems, not four.
Every additional stem is another category the model must sort every sound into. A two-way split asks one question: voice, or not voice? A four-way split asks three, and each decision is another place for a mistake.
The "other" stem in a four-way split is also where the uncertainty collects — it holds whatever did not confidently belong anywhere else, which is why it often sounds the muddiest.
Start from the best file you have
A lossy file has already had detail removed by its encoder, chosen on the assumption that a human listening to the whole mix would not notice. A separator is not listening to the whole mix — it is examining the parts, including exactly the quiet detail the encoder judged disposable.
That damage then gets distributed across the stems, and no setting recovers it.
So: WAV or FLAC where you have the option. A high-bitrate MP3 is usually acceptable. A heavily compressed file, or one that has been re-encoded more than once, will produce a poor separation no matter which tool you use, and swapping apps will not fix it.
What "good enough" means for your use
This is the part worth deciding before you judge a result.
Singing along at home. A faint trace of the original in the chorus is irrelevant — you are singing over it. Almost any separation is good enough.
Practising an instrument. You need the part you are learning gone and the rest audible. Minor artifacts are fine; what matters is isolation of the thing you are replacing.
Backing track for a live performance. Higher bar. Artifacts you cannot hear on headphones can become obvious through a PA in a quiet room.
Release or distribution. Separation is the wrong tool. Beyond the audible quality, the copyright in the recording is unchanged by the fact that software removed a layer from it — separating a track you own for private practice is a different matter from publishing or performing the result, and if that is the plan, check what applies where you are before you rely on it.
Most disappointment with these tools comes from judging a home-practice result by a release standard.
A note on the app I build
Disclosed interest: Vocly is mine.
It does two stems or four, with a standard and an HQ model so you can trade speed against quality, and everything runs on the phone — the audio is not uploaded. The features that matter most given everything above are the practice ones: mute and solo stems in real time, loop a section, slow the tempo between 75 and 125 percent, shift pitch by up to six semitones, and A/B against the original, which is the honest way to hear what the separation actually did.
It will not give you a perfect instrumental on a dense live recording with layered harmonies. Nothing will. What it can do is give you the best estimate available for that song, on your device, and let you balance the stems yourself rather than accepting one fixed result.
Frequently asked questions
Why is there still a trace of vocals in my instrumental?
Because the separator never had the vocal track. It only had the finished mix, and it estimated which parts of that mix were voice. Where the voice overlaps something else in frequency and time — or where reverb has smeared it across the stereo field — the estimate is imperfect, and what is left over is what you hear.
Which songs separate well and which do not?
Sparse, dry, modern recordings with one clear lead voice separate well. Dense mixes, heavy reverb, doubled or layered vocals, choirs, live recordings and loud distorted guitar are hard, because the voice and everything else occupy the same space.
Should I use two stems or four?
Two, if you only need vocals and instrumental. Every extra stem is another decision the model has to make about every sound, so four-way separation generally has more bleed than a two-way split of the same song.
Does the source file quality matter?
Yes. A lossy MP3 has already had detail discarded by the encoder, and the separator cannot recover what is not there — it works on the damaged version and the damage ends up distributed across the stems. Start from WAV or FLAC when you have the choice.
Can I legally make a karaoke version of a song?
Separating a track you own a copy of, for your own practice at home, is a different situation from publishing or performing the result. Copyright in the recording does not change because software removed a layer from it. If you intend to upload or perform it, check the rules that apply where you are.