Why can I still hear drums in the vocal stem?

Because separation is an estimate, not an undo. The parts overlap in frequency and in time, and where two sources occupy the same space the split has to guess which one owns it. Cymbals and vocal sibilance are the classic pair — both are broadband, both sit high, and they are genuinely hard to tell apart.

Does the file format I upload matter?

Yes, more than any setting in the tool. Lossy encoding removes the fine high-frequency detail the separation uses to tell sources apart, and it removes it permanently. A lossless file gives the best result, a high-bitrate MP3 is usually fine, and a file that has been encoded more than once — downloaded from a re-upload, or lifted from a video — is where results fall off sharply.

Why is my instrumental quieter than the original?

Because it is. Removing a part removes energy from the mix, and the loudness a mastered track carries came partly from everything sitting together. A stem sounding thinner and quieter in isolation is normal and not a fault — level it back up afterwards rather than reading it as a failed split.

Would a deeper split give me a cleaner vocal?

No. The vocal is separated from the rest either way, and asking for more parts does not sharpen it — it just divides the remainder further. If a Band split gave you a bled vocal, a Detailed split will give you the same bled vocal plus more files.

What kinds of recording separate worst?

Live recordings with crowd noise, heavily distorted or wall-of-sound mixes, anything mastered extremely loud, mono recordings, and old or noisy transfers. All of them share one property: the sources were never cleanly distinguishable in the file to begin with.

Can I re-run a split with different settings?

Yes, and it is a new job charged at the same rate — the tool does not keep a half-finished result to refine. That is a reason to think for ten seconds about the mode before running, and a reason to try the cheaper shallower split first when you are unsure.

Stem Splitter

Getting a clean separation

The recording you feed in decides most of the result. What bleeds and why, why an MP3 of an MP3 is the worst case, and when a six-part split is worse than a four.

Lesson 3 · core · 7 min read

Open Stem Splitter

It is tempting to treat separation as an undo button for mixing. It is not. A finished song is one waveform, and the parts were added together in a way that cannot be reversed — what a separator does is estimate what each part probably sounded like, from what it can still hear of it. Everything in this lesson follows from that one fact.

The source decides the result

The largest single factor in how a split turns out is not a setting. It is the file you uploaded. Lossy compression works by discarding detail your ears are unlikely to miss — and a good deal of that discarded detail is exactly what the separation was going to use to tell two sources apart.

Write this

Not this

Where the file came from

The original file you bought, ripped, or exported yourself

Ripped from a video, or downloaded from a re-upload of a re-upload

Encoding history

Encoded once, from the master, or never encoded at all

Encoded, decoded and re-encoded several times over

The recording itself

A studio mix with the instruments spread across the stereo field

A live take with crowd noise, or a mono transfer of an old record

The master

Dynamic, with space between the loud parts and the quiet ones

Crushed as loud as it will go, everything squashed together

Where the file came from

Write this:: The original file you bought, ripped, or exported yourself
Not this:: Ripped from a video, or downloaded from a re-upload of a re-upload

Encoding history

Write this:: Encoded once, from the master, or never encoded at all
Not this:: Encoded, decoded and re-encoded several times over

The recording itself

Write this:: A studio mix with the instruments spread across the stereo field
Not this:: A live take with crowd noise, or a mono transfer of an old record

The master

Write this:: Dynamic, with space between the loud parts and the quiet ones
Not this:: Crushed as loud as it will go, everything squashed together

Careful

An MP3 of an MP3 is the worst realistic case. Each encode throws away a different slice of the top end, and the second pass throws away detail that was already damaged. You cannot get it back, and no mode or preset in this tool will make up for it — so if there is any chance of finding a cleaner copy, that is a better use of five minutes than re-running the split.

What bleeds, and why

Bleed — hearing a bit of one part inside another — is not a bug that will be fixed in a later version. It is what happens where two sources genuinely occupy the same frequencies at the same moment, and the separation has to attribute that shared energy to one of them.

Where you hear it: Why it happens
Cymbals in the vocal: Both are broad, bright and noisy rather than pitched. Sibilance and a crash share the same region, and there is often no clean line between them.
Kick drum in the bass: They live in the same low octaves and usually play on the same beats. A bass note under a kick hit is one blob of low-frequency energy.
Reverb and room in the wrong part: Reverb was printed into the mix as part of the sound. The tail of a vocal reverb is not clearly a vocal, so it tends to land in whatever part catches the leftovers.
Doubled or harmonised vocals: Backing lines that move with the lead share its timbre and timing. A separator splitting lead from backing is drawing a line the recording does not clearly contain.
Distorted guitar smeared everywhere: Distortion spreads a single instrument across the whole spectrum, so it overlaps almost everything else at once.

The pairs that bleed most often, and what they have in common.

Note

Almost every split has one weak file, and it is usually the one that catches whatever is left over. That is not the split failing — it is the leftovers going somewhere. Judge the job on the part you actually needed.

When six parts is worse than four

More parts sounds strictly better and is not, for a reason worth stating plainly: a part that is not in the recording does not come back empty. It comes back as a file with something else in it.

Run a Detailed split on a guitar-band song with no piano and you get a piano file. It will contain something — sustained notes from a synth pad, the harmonic body of a guitar chord, whatever most resembled a piano. Meanwhile those sounds have been taken OUT of the part they belonged in, so you have not just gained a useless file, you have degraded a useful one.

Ask what is actually playing before choosing a depth. Band for most songs; Detailed when there really is a distinct piano and guitar to find.
Karaoke is the cleanest split available, because two-way separation has the least to get wrong. If all you need is “with vocal” and “without”, take it.
The same trap applies to naming instruments: picking a saxophone from the catalogue when the song has no saxophone gives you a saxophone file, not an empty one.
If a part bled, go shallower rather than deeper. Bleed means you asked for a distinction the recording does not support, and more parts asks for more of them.

Tip

When you are unsure, run the shallower split first. It costs exactly the same as the deep one, because you are billed for the length of the song rather than the number of parts — so the only thing you lose by starting simple is a minute, and what you gain is a clean baseline to judge the deeper attempt against.

What a good result sounds like

Set the expectation correctly and you will stop re-running jobs that already worked. A good stem is usable, not pristine. It will be quieter and thinner than the same part sounded inside the mix, because the loudness of a master comes partly from everything sitting together. It will have some faint bleed at the edges. It may sound slightly watery under headphones on the quietest passages.

None of that is a failed split. What a failed split sounds like is different and unmistakable: the part you asked for is largely absent, or another instrument is clearly present at full volume. That is the result worth re-running — usually with a better source file, or a shallower split.

Questions

What people ask about this

Free to start

Now go make one

Reading about a style description only gets you so far. The studio is free to use — write one sentence and hear what comes back.

Create a free account

1,000 credits every month. No card required.

All Stem Splitter lessons

The source decides the result

Write this

Not this

Where the file came from

The original file you bought, ripped, or exported yourself

Ripped from a video, or downloaded from a re-upload of a re-upload

Encoding history

Encoded once, from the master, or never encoded at all

Encoded, decoded and re-encoded several times over

The recording itself

A studio mix with the instruments spread across the stereo field

A live take with crowd noise, or a mono transfer of an old record

The master

Dynamic, with space between the loud parts and the quiet ones

Crushed as loud as it will go, everything squashed together

Where the file came from

Write this:: The original file you bought, ripped, or exported yourself
Not this:: Ripped from a video, or downloaded from a re-upload of a re-upload

Encoding history

Write this:: Encoded once, from the master, or never encoded at all
Not this:: Encoded, decoded and re-encoded several times over

The recording itself

Write this:: A studio mix with the instruments spread across the stereo field
Not this:: A live take with crowd noise, or a mono transfer of an old record

The master

Write this:: Dynamic, with space between the loud parts and the quiet ones
Not this:: Crushed as loud as it will go, everything squashed together

Careful

What bleeds, and why

Where you hear it: Why it happens
Cymbals in the vocal: Both are broad, bright and noisy rather than pitched. Sibilance and a crash share the same region, and there is often no clean line between them.
Kick drum in the bass: They live in the same low octaves and usually play on the same beats. A bass note under a kick hit is one blob of low-frequency energy.
Reverb and room in the wrong part: Reverb was printed into the mix as part of the sound. The tail of a vocal reverb is not clearly a vocal, so it tends to land in whatever part catches the leftovers.
Doubled or harmonised vocals: Backing lines that move with the lead share its timbre and timing. A separator splitting lead from backing is drawing a line the recording does not clearly contain.
Distorted guitar smeared everywhere: Distortion spreads a single instrument across the whole spectrum, so it overlaps almost everything else at once.

The pairs that bleed most often, and what they have in common.

Note

When six parts is worse than four

More parts sounds strictly better and is not, for a reason worth stating plainly: a part that is not in the recording does not come back empty. It comes back as a file with something else in it.

Ask what is actually playing before choosing a depth. Band for most songs; Detailed when there really is a distinct piano and guitar to find.

Karaoke is the cleanest split available, because two-way separation has the least to get wrong. If all you need is “with vocal” and “without”, take it.

The same trap applies to naming instruments: picking a saxophone from the catalogue when the song has no saxophone gives you a saxophone file, not an empty one.

If a part bled, go shallower rather than deeper. Bleed means you asked for a distinction the recording does not support, and more parts asks for more of them.

Tip

What a good result sounds like