Stem Splitter

A stem splitter is a program that takes a finished piece of music apart. You feed in a song and get out several individual tracks: vocals on one track, drums on another, bass on a third. In the music studio, these individual tracks are called “stems”. In the past, once a song had been mixed down, it couldn’t be taken apart again — it was like trying to turn scrambled eggs back into a raw egg. Modern stem splitters manage it anyway, because they’ve been trained on huge amounts of music data and have learned which sound patterns typically belong to which instrument.

Significance for musicians, producers, and fans

For professional music production, source separation was long an expensive niche topic. Anyone wanting to isolate an instrument from someone else’s song needed either access to the studio’s original files or a great deal of patience for manual editing. Neither option is realistic for most people.

Stem splitters democratize this access. A hobby musician can now isolate the drums from a favorite song to use as a practice backing track. A DJ can cut out the a cappella vocals of a track and lay them over a different beat. And music teachers can mute individual instruments so students can listen closely to exactly what the rest of the ensemble is playing.

Related use cases also exist in speech processing: systems that separate individual speakers in a recording with multiple voices work on the same basic principle. However, the term “stem splitter” almost exclusively refers to music.

How an AI model separates sounds

An audio signal is a single waveform — all the instruments overlap within it simultaneously. The model doesn’t learn to mechanically “untangle” this waveform. Instead, the recording is first converted into a different representation: a so-called spectrogram, which shows which frequencies are how loud at which point in time. Low frequencies at the bottom, high frequencies at the top — similar to a heatmap.

The model — in practice often a neural network, i.e. a system made up of many mathematical layers — was trained on thousands of songs for which the original tracks were available. In the process, it learned which frequency patterns typically come from a voice and which come from a bass guitar. When applied to a new, unknown song, it estimates for every point in time and every frequency which instrument the sound most likely belongs to. From this, it reconstructs the separated tracks.

The result is rarely perfect. High female voices and bright guitars sit close together in the frequency spectrum. When both sound at the same time, errors occur: a bit of guitar ends up in the vocal track, or vice versa. This artifact is called “bleeding” — meaning the signal “bleeds” into the wrong track. Quality depends heavily on how many and how varied the pieces of music the model has previously seen.

Stem splitters in practice: tools, legal disputes, and remixes

The best-known freely accessible tool is called Spleeter, developed by the music streaming service Deezer and released as an open-source project in 2019. Commercial alternatives such as LALAL.AI, Moises, or iZotope RX target professionals and promise better quality on difficult recordings. DJ software such as rekordbox or Traktor now has stem splitting built in directly.

The technology also sparks debate, however. Anyone who isolates an artist’s vocals from a song and uses them in a new context is operating in a copyright gray area. Music publishers have already examined cases in which isolated vocals were used for unauthorized remixes. The software itself is legal — what someone does with the result is a different question.

In pop culture, stem splitting became more widely known when, after the deaths of musicians like Michael Jackson or Prince, fans began extracting and releasing a cappella tracks from old albums. This sparked both enthusiasm and legal disputes at the same time — and showed just how much this technology is changing the relationship between music, audiences, and copyright.