Guides · checked 1 Nov 2026
Stems from the session, or separated stems
Two files can both be called a stem. One was printed from the producer's session. The other was guessed from a finished mix by a model. For training data the difference is the whole point.
How a session stem is made
The producer opens the project, and the DAW prints each track to its own file with its plugins running. A song of 26 tracks gives 26 files on the same clock, each one instrument or one layer, named by the person who made it. Summed at unity gain they give the mix, within rounding. There is no bleed, because nothing was ever mixed together to begin with. Transients, reverb tails and the top octave are all intact.
How a separated stem is made
A separation model takes the finished mix and estimates, for four to six sources, which part of each moment belongs to which source. It is a guess, and a good one on a clean pop mix. The guess still leaves bleed between sources, smeared transients, holes in the spectrum where two instruments overlapped, and a residual that does not cancel. Everything that was never a source of its own, a riser, a vocal chop, a sub layer, lands in "other".
Why it matters for training
A model trained on separated stems cannot get better than the separator that made them. It learns the bleed and the holes as truth. Scoring it against separated "ground truth" is circular. For mixing, mastering and generative models the loss is larger: a separated stem has no plugin chain, no MIDI, no lane name and no relation to a session, so there is nothing to learn about how the sound was made.
How to tell them apart
- Null test: sum the stems and subtract the mix. Session stems cancel to near silence. Separated stems leave a residual you can hear.
- Count: four to six files per song means a separator. Session stems come in the dozens, with names like "Pluck 2" and "Kick top".
- Company: session stems arrive with the project, the MIDI, the presets. Separated stems arrive alone.
- Look: a spectrogram of a separated stem has soft holes where another instrument sat. A printed lane has a floor of silence.
- Paper: ask the vendor to warrant in the licence that no file was made by a separation model.
What XIM Audio holds
4,223 songs, 110,662 lanes printed from the producers' own sessions, about 9,153 hours, measured 4 Nov 2026. One 24-bit WAV per lane with one of 53 instrument labels, taken from the producer's lane name. The unmastered mix of each song ships as the unity-gain sum of its lanes, 32-bit float, with the residual of a second sum recorded per song. Bleed is false by construction. 3,571 songs come as paired sets with the session file itself, and 9,468 DAW sessions stand on their own. Our standard licence warrants that no file was made by a separation model or by AI. See the stems set.