A finished song combines several sources into a small number of audio channels. Recovering the vocals, drums and other instruments is not the reverse of opening a project file. Much of the information about their original separation is absent from the mix.
Deezer's Spleeter turns that difficult source-separation problem into an accessible tool with pretrained models. The initial documentation describes two-, four- and five-stem configurations, ranging from vocals plus accompaniment to a more detailed grouping of instruments.
The output categories are useful estimates. They are not guaranteed copies of the original studio tracks. A vocal shares frequencies with guitars, keyboards and other sounds. Reverb spreads a source across time, and distortion changes its spectral shape. The model has to infer a plausible allocation from patterns learned during training.
A spectrogram provides a useful way to picture the task. Time runs in one direction, frequency in the other, and intensity describes energy. Musical sources create patterns in that space. A separation model can estimate how much of the mixture belongs to each source, but several sources may overlap in the same region. There is no universal frequency boundary above which vocals live and below which accompaniment begins.
That is why listening to an isolated stem and listening to it in a new mix are different tests. A small remnant of an instrument may be obvious in a solo vocal but masked when the vocal is placed over other material. Conversely, an artifact that seems minor alone may become distracting after compression or time stretching.
I would evaluate a separation with a specific creative task in mind. For studying a bass line, intelligible notes may matter more than a perfectly clean timbre. For preparing an exposed vocal passage, small watery or metallic artifacts may be unacceptable. There is no single quality threshold that serves every use.
A useful comparison also includes the original mix at a matched listening level. Louder output can seem clearer, and a missing component can make a stem feel cleaner while removing part of the intended source. Short difficult passages are often more informative than one favorable section repeated several times.
What makes Spleeter interesting is the reduction in setup between a research idea and a concrete question about a piece of audio. A musician or developer can examine a result without first training a model from scratch. The remaining judgment is still human: whether the estimate is good enough for the intended use, and which artifacts matter in that context.
The tool gives us a new way to inspect a mix. It doesn't turn the mix back into its original session, but it can make musical structure much easier to hear.