Stem Separation
Stem separation that keeps the tail of the sound
Separate a stereo mix into vocals, drums, bass and the rest — or a five-way split with piano — at the full sample rate you gave it.
Meters at ×1.5 — a four-minute track costs 6 minutes of your allowance.
How it works
What it measures, what it changes
And what stays yours to decide. Every stage is a parameter you can read back and overrule.
- 01
The mix is analysed before it is split
Tempo, key, the transient map and which frequency regions actually contain each source are established first. A separator that starts from the spectrogram alone has to work all of that out in the same pass that's doing the splitting.
- 02
You choose how many ways it splits
Two stems (vocal and instrumental), four (vocals, drums, bass, other) or five (adding piano). More stems isn't automatically better — each extra head has to take its material from somewhere, and a four-way split of a track with no piano comes out cleaner than a five-way one.
- 03
The separation runs at full rate
The model works on the source at its own sample rate instead of downmixing to 44.1 kHz and upsampling the result, so what comes back isn't band-limited by the separation step. The stereo image is preserved, not reconstructed from a mono estimate.
- 04
The stems stay in the project
Separated stems become part of the track's project, not a folder of downloads. Tune the vocal you just extracted, or master the mix you rebuilt from the stems — each stage knows what the previous one did.
The specifics
What you would ask before trusting it with a record
Formats, limits, and the places it is honestly hardest. Nothing here is a range we have not measured.
- Stem counts
- 2-stem (vocals / instrumental), 4-stem (vocals / drums / bass / other) and 5-stem (adding piano). Each is its own head, not a recombination of one split after the fact.
- Sample rate and depth
- Stems come back at the source sample rate up to 192 kHz, 24-bit or 32-bit float. No forced downmix to 44.1 kHz, and no upsampling a lower-rate result to disguise one.
- Sum behaviour
- The stems sum back to the input within the residual the model leaves, and the residual itself comes back inside the 'other' stem rather than being thrown away. Nothing in your track quietly disappears between the input and the sum of the outputs.
- Where it is hardest
- Dense low-mid material, heavy bus compression on the source, reverb tails that belong to a source being pulled out, and doubled vocals panned wide. Those are the cases we measure the quality bar on, not the easy ones.
- What it is not
- Separation recovers what's present in the mix; it can't recover what was never there. A stem pulled from a stereo bounce is not the multitrack.
Questions
What people actually ask
- How many stems should I ask for?
- The fewest that answer your question. If you want an instrumental, ask for two — the four-way split has three more chances to leak. If you're rebuilding a mix, four. Five only if the track genuinely has a piano to find.
- Why does the vocal stem have reverb in it?
- Because the reverb is part of the vocal in the mix. A tail that was printed onto the vocal bus belongs to it acoustically. Pulling it off is a different job — dereverb, over in restoration — not something separation can decide for you.
- Can I use the stems commercially?
- That depends on your rights to the recording, not on us. Separating a track you don't have the rights to process doesn't create rights in the result. Our acceptable-use policy spells this out, and the warranty is the uploader's.
- Why does this cost more per minute than mastering?
- Because it takes materially more compute. Every capability's multiplier is published — separation meters at 1.5× the base minute rate — so the difference is a number you see before you submit, not a surprise on the invoice.
- Is this the same thing as a stem splitter?
- Yes. Stem splitter, song splitter, audio splitter and stem separation all name the same job. This is Euphona's version of it: the mix analysed first, then split into music stems at the sample rate you gave it.
Try it
Put your own track through it
The free plan runs the whole engine on your whole song and plays you the result, matched for loudness against the original. No card. You pay when you want the file — one track at a time, or a plan if you are making a lot of them.
Who it’s for