How Does AI Stem Separation Work?

AI stem separation works by training a neural network on thousands of songs where the individual parts are known, so it learns what vocals, drums, bass and other instruments look like inside a mixed signal. Given a new song, it predicts how much of each sound belongs to each stem and writes out four separate files.

The training idea

A mix is the sum of its parts. If you have the parts, you can make the mix, but you cannot normally go backwards, because the information about which instrument made which part of the waveform is gone. Separation models learn that mapping statistically.

Training data comes from multitrack recordings: a song plus its isolated vocals, drums, bass and other. The model is shown the mix and asked to produce the stems; the difference between its guess and the real stems is the error, and the network's weights are adjusted to reduce it. After enough songs the model has internalised what a snare transient, a sung vowel or a sub bass fundamental looks like in time and frequency.

How a song is processed

1. The audio is cut into short overlapping chunks, a few seconds each.
2. Each chunk is analysed both as a waveform and as a spectrogram, a picture of frequency over time. Hybrid models such as Demucs run both views in parallel.
3. For every point in the spectrogram the model estimates a mask: what fraction of that energy is vocals, drums, bass and other.
4. The masks are applied, the chunks are turned back into audio and stitched together.
5. The result is four files of identical length that sum back to something very close to the original mix.

Why "other" exists

Guitars, keys, synths, strings and pads do not have a reliable fingerprint the way drums and voices do, so most models group them together as "other". Newer models split out guitar and piano, but for DJ work the four-stem split is standard and is what Club Ready Music uses.

What it gets right and wrong

Separation is best on dense modern productions with a clear vocal, steady drums and a bass that lives below 150 Hz. It struggles with:

• Heavily distorted or chorused vocals that look like synths.
• Drums with long reverb tails, which can land in "other".
• Bass guitar played high up the neck, which overlaps with keys.
• Live recordings with room bleed.

Artifacts are mostly a slight watery or warbly quality on isolated stems. When stems are summed back together, as in a DJ edit, the artifacts cancel and the result sounds like the original.

How Club Ready Music uses the result

In DJ Edit, the uploaded song is sent to a Demucs model that returns vocals, drums, bass and other as WAV. Club Ready Music searches the drum stem for the loudest four-bar window and loops it for the intro; fades the vocal and other stems over the outro while drums and bass continue; and sums everything back for mastering. The stems are not exposed to the user on DJ edits. In the Studio, generated tracks can include stems as a download.