Search for "remove vocals from a song" and you will find two kinds of tools: vocal removers and stem splitters. The names sound like two different technologies. They are not. Both use the same family of neural networks trained on multitrack recordings; the difference is in what they output and what they are tuned for. Knowing which one you need saves processing time and gives you a better result.
The short version
- A vocal remover gives you two files: the voice and everything else. It is tuned to make the "everything else" as clean as possible. Use it for karaoke, backing tracks and acapellas.
- A stem splitter gives you four or five files: vocals, drums, bass, piano and "other" instruments. Use it for remixing, practice tracks where you mute your own instrument, and transcription.
Many services, including the Mazmazika vocal remover, offer both modes from the same page, so the real question is which mode to choose for the job.
How both of them work
A finished song is a single stereo file. Nothing in it says "this part is the singer". A source-separation model learns, from thousands of songs where the individual tracks were available, what voices, drums, basses and keyboards tend to sound like: their timbre, their position in the stereo field, the way they move over time. Given a new mix, it predicts, for every small slice of time and frequency, how much of the energy belongs to each source, and rebuilds one audio file per source from those predictions.
That is why the results are never quite perfect. Where two sources overlap in time and frequency, for instance a female voice and a distorted guitar in the same register, the model has to make a judgement, and a little of one ends up in the other. Larger models make better judgements; that is what the "High Quality" and "Ultra" engines are.
For a plain-English walk through the model itself, see How AI vocal separation actually works.
Why two stems can be cleaner than five
It is tempting to think that five stems is simply "more". For the instrumental it is often worse. When a model splits five ways it has to assign every bit of energy to one of five buckets, and errors accumulate; a snare hit that lands partly in "other" and partly in "drums" sounds slightly smeared when the two are played together again. When it only has to decide "voice or not voice", there is one boundary to get right, and the instrumental keeps the original mix's glue.
So, for karaoke and backing tracks, choose two stems. For anything where you actually want the drums or bass on their own, choose the multi-stem mode and accept a little more artefact per stem.
Which one for which job
Karaoke and singing practice. Vocal remover, two stems. Keep the harmonies by using a lead-vocal-only model if the tool offers one.
Acapella for a remix or mashup. Vocal remover, two stems, and use the highest-quality engine you have; the acapella is the stem you will hear most exposed.
Practice tracks for a band member. Stem splitter. Mute the bass stem to practise bass, the drums stem to practise drums, and so on. The browser mixer does this without a DAW.
Transcribing a part. Stem splitter. Soloing the piano or bass makes the notes far easier to hear, and running the result through the chord analyzer gives you the harmony to check against.
Sampling a drum break. Stem splitter, drums stem. Expect a little bleed from other percussive elements.
Remixing a whole song. Stem splitter, five stems, then treat each stem as a track in your DAW. Re-EQ and compress each one lightly; separated stems respond well to that and it hides the seams.
Quality: what to expect from each
Both modes are strongest on modern, well-produced material and weakest on dense, reverberant or lo-fi recordings. A few specifics:
- The instrumental from two-stem mode is usually the cleanest single output any separation tool produces.
- The acapella carries the most artefacts, because reverb and effects that were on the voice are partly ambiguous.
- The drums stem is generally excellent; drums are spectrally distinct.
- The bass stem is good but can swap notes with a kick drum's low end in very bass-heavy mixes.
- "Other" is the catch-all and the least predictable: guitars, synths, strings and anything unusual land here together.
Free or paid?
Two-stem separation with the standard engine is free on Mazmazika, with a per-run track length and a daily allowance. Which of the larger engines, models and track lengths are included in the free plan and which come with Pro is listed live on the vocal remover page under Free vs Pro, and the pricing page shows current prices in your currency. If you are making one karaoke track for a party, the free path is enough; for stems you will release or play on stage, the larger engines pay for themselves in the first track.
A note on the old methods
Before neural separation there were two tricks: phase cancellation, which subtracts the left and right channels to remove anything panned to the centre, and EQ notching, which carves out the vocal frequency range. Both damage the whole mix, because the bass, kick and snare are also in the centre, and the vocal range is where guitars and keys live too. They are still built into some DAWs and free plug-ins; there is no reason to use them today.
Next steps
- Try both modes on the same song with the vocal remover and stem splitter and compare the instrumental.
- Making a karaoke version? Follow the step-by-step karaoke guide.
- Curious about the model? Read how AI vocal separation works.