How stem separation actually works
Stem separation looks like magic and gets described like filtering. It is neither. Understanding roughly what it does explains most of the quality problems people hit.
It is reconstruction, not subtraction
Old vocal-removal tricks worked by phase cancellation: if the vocal is centred and the rest is not, invert one channel against the other and the centre disappears. That takes out the bass and the kick too, and it only works on one specific kind of mix.
Modern separation is a trained model — Demucs is the one dhun uses — that has learned what a drum kit, a bass, a voice and "everything else" look like in the time-frequency domain. Given a mix, it predicts each source separately and outputs it. It is not removing things from your file; it is generating four new files that add back up to approximately the original.
Why that explains the artefacts
Because each stem is a prediction, the failure mode is not "some of the vocal is left behind" but "the model was unsure". Where two sources overlap heavily in frequency and time — a snare and a hi-hat, a bass and a kick drum, a vocal and a saturated guitar in the same range — the model has less to go on, and you hear that as smearing or a faint watery quality.
This is also why separating a sparse, well-recorded mix gives near-perfect stems while separating a loud, heavily limited master gives worse ones. Limiting reduces the dynamic differences the model uses to tell sources apart.
If you have the option, separate from the least-processed version you have. A pre-master mix separates noticeably better than the loud master of the same song.
What it is genuinely good for
Getting an acapella or an instrumental from something you only have as a stereo file. Remixing. Rescuing a part from a session whose project file is long gone. Practising along with a track minus one instrument. Sampling responsibly, where you have the rights.
What it is not good for is treating a separated stem as if it were the original multitrack. A separated vocal has been through a reconstruction; it will not take the same aggressive processing an original take will.
Doing it in dhun
Upload the track and ask it to split the stems. It runs server-side, which is why the phone app can do it as well as the desktop — the work is not happening in your browser.
Stem separation is unlimited on every paid plan. It costs CPU rather than vendor money, so there is no reason to meter it.