Vocal Separation vs Stem Separation: What Each One Is For | WaveCraft

The short answer

Vocal separation is a two-way split: you get the vocal and everything else, and that decision is made for you. Stem separation is a four-way split: vocals, drums, bass and the remaining harmony, each one a track you can edit on its own.

They are the same technique at two settings, so the question is not which is more advanced. It is what you plan to do with the output. If you want a backing track to sing over, two ways is enough and four is wasted effort. If you want to change one instrument and leave the rest alone, only the four-way split gets you there.

WaveCraft runs both from the same on-device engine. It is on the App Store — the rest of this page is about which one your track needs.

Why you are asking this

You have a track you like, and you want something out of it that it does not currently contain. The versions of that want are specific, and they need different outputs:

• "I want to sing over this." You need the accompaniment. A two-way split produces it in one pass.
• "I want to use just the drums." You need the drums without the rest competing for your attention. That output exists in the two-way result too.
• "I want to change the song." Replace the beat, rebalance the low end, strip the vocal and put your own in its place. This is the only one that genuinely needs all four parts as independent tracks.

Most people land in the first group and buy a tool built for the third. That mismatch is the whole reason the question comes up: "stems" is the industry's word for the most powerful mode, and it is not the mode most people need.

What the two outputs actually are

Vocal separation produces two tracks that sum back to the original: an estimate of the lead vocal, and everything the model judged was not the lead vocal. That second track is usually called the instrumental, and it is a reconstruction rather than an extraction. Every backing track described as "the instrumental version" is this output.

Stem separation produces four. Vocals and drums and bass are the named ones; the fourth is usually called "other" or "harmony", and it is whatever the model could not confidently name — piano, guitars, strings, pads, sound effects, and often the bleed that belongs to the other three. That fourth stem is where the technique's honest limits show up.

The two are not two qualities of the same thing. They are two different questions asked of the model: is this the singer? versus which instrument is this? The first is a much easier question than the second, and everything else follows from that.

Why four-way separation can sound worse than two-way

This is the counter-intuitive part, and it is the reason "more stems" is not automatically better.

A two-way separation has one bucket for anything that is not the singer, and nothing in that bucket needs to be identified. A four-way split has to place every part somewhere. Four names — vocal, drums, bass, harmony — for a dense modern mix with more than four sound sources competing. Whatever the model is unsure about goes into the leftover harmony stem, which is then not one instrument but a mixture. Sorting out what to do with it becomes your problem.

What that looks like in practice:

• The drums arrive with vocal reverb on them. In a four-way split, the model's "drums" output carries a little of the vocal's tail, because the two overlap in time and frequency at every chorus.
• The bass and the kick share a stem. In nearly every mix those two live in the same octaves. Separating one from the other is not a clean spectral cut, so whichever the model is less sure about partly lands in the harmony bucket.
• The "harmony" stem is an instrument only at the start of the track. In a song that builds, the new parts in the bridge accumulate in the leftover as the track goes on.

Why the cheap version fails

Before trusting any output, it is worth knowing what a bad one sounds like, because the failure is usually a technique rather than a quality setting.

The common shortcut assumes the singer is panned dead centre, then subtracts one stereo channel from the other. On a 1960s stereo recording with the vocalist exactly centre-panned, this works surprisingly well. On anything modern it fails, because the centre of a modern mix is full of everything: kick, bass, snare, piano, and sometimes a second singer.

What comes back is a thin, phasey version of the mix with the bass and part of the vocal missing too. It is not a setting you can fix. The information was never separated; it was averaged.

A trained separation model instead estimates a mask — a per-frequency, per-moment estimate of how much of the signal is the singer — and reconstructs both signals from it rather than subtracting. It never assumes the singer owns the middle.

What each one costs you

Separation is estimation, so quality depends on the source. Honest expectations:

Both outputs survive these: lead vocals sitting clearly above the mix, dense mastered pop, most studio recordings, anything with moderate reverb on the vocal.

Both degrade on these: heavy vocal distortion or screaming, several singers overlapping, extremely compressed masters where the vocal and the instruments occupy the same frequencies, and very old or very lo-fi recordings where the model has little to learn from.

The four-way split additionally suffers when: the arrangement is dense, the track has no clear low-end division between kick and bass, or the song changes instrumentation halfway through.

Two more limits worth stating plainly. Separation does not improve the source. The separated vocal is a reconstruction of the same performance — not a re-recording, and never cleaner than what was recorded. And separation is not a licensing tool. Posting a separated instrumental publicly is still a copy of someone else's recording, and content-ID matches the recording, not the vocal.

The property that makes stems worth having

When a separation is complementary — every output summing back to exactly the original mix — separation is reversible rather than lossy. Recombine the parts in any configuration and the original is recoverable.

That matters more for stems than for a two-way split, because stems invite you to experiment. A four-way split you cannot undo means every re-mix is a one-way door: layer the parts, hear that the mix came out thinner than the source, and there is no way back except re-separating.

If you export stems from any tool and find that stacking them produces a mix permanently quieter or duller than the original, that is a property of the tool, not of your project.

What WaveCraft does differently

The engine is the same for both modes, and it runs on your device — the audio is never uploaded and the pass works with no connection at all. What matters is what the split feeds into:

• Unlimited tracks, so a four-way split arrives as four tracks you can mute, solo, time-stretch and arrange rather than as four files to import.
• A real timeline — split, trim, fade, blend, time-stretch and per-track volume, so nudging a stem a few milliseconds against the others is a two-second adjustment rather than a re-export.
• 20+ effects and mastering presets — EQ, reverb, compression, limiting and modulation. Denoising a separated vocal or EQing the bass stem alone is the point of splitting in the first place, and it needs to happen in the same project.
• Auto-mix and loudness normalization, so four separate tracks can be brought back to a sensible balance.
• On-device AI for separation, denoise and mixing, in seconds.
• Lossless export up to 32-bit float in WAV, ALAC, FLAC or AAC.

The honest limit is the model itself: a four-way split of a dense arrangement will still hand you a harmony stem that is a mixture rather than an instrument, and no setting changes that. Pro unlocks unlimited separation passes, the full effects set and lossless export; the free tier lets you try the core editor, with a 7-day trial on the subscription plans.

The decision

Start with a two-way vocal split. It is the cheaper operation, it is the one that sounds more finished on most tracks, and if you are singing, looping, learning a part or making an a cappella, it is the end of the job. Get the accompaniment, check it for bleed, and see whether you still want more.

Move to stems when you have a specific instrument to change. The test is whether you can name the part: replacing the drums, reworking the low end, looping one element, or putting your own vocal over the original. If your plan is more general than that, you are paying for four-way separation to get the same two files you already had.

WaveCraft runs both from one on-device engine either way, so the choice costs you nothing but the processing time. Download it on the App Store — or read how to remove vocals from a song if you are starting from the two-way version, and how to make a karaoke track for the licensing question that comes up once the result is good enough to publish.