Stem separation is the process of taking a finished, mixed song and pulling it apart into its parts, such as vocals, drums, bass and the rest. Today it is done with AI models trained to recognize each kind of sound. You can try it on your own song with the Unmix Studio stem splitter, which returns six parts as MP3 files.
If you want to know what a stem is first, read what are stems in music. This article is about how the separation works.
When a song is mixed, every instrument is added together into one waveform. The voice, the guitar and the snare all move the same speaker at the same time. There is no hidden layer that still holds each one.
Think of it like a smoothie. The fruit went in separately, but you can't simply pour the strawberries back out. To get them back, you need something that knows what a strawberry tastes like.
Before AI, people used a trick called phase cancellation. In many songs the lead vocal sits in the center, equally loud in the left and right speakers. Flip one side upside down and add it to the other, and anything in the center cancels out.
It worked a little. But the bass and kick drum are usually in the center too, so they vanished with the voice. Reverb on the vocal, which spreads to both sides, stayed behind. The result often sounded thin and hollow.
Modern stem separation uses neural networks. A model is shown thousands of songs together with their real stems. Over time it learns what a voice looks like in sound, and what a drum or a bass looks like.
When you upload a song, the model looks at the sound over time and frequency. For each tiny slice it estimates how much belongs to each part. Then it rebuilds each part as its own audio file.
Because the model learned the sound of each instrument, it can separate a bass from a voice even when both sit in the center.
A typical split gives you a fixed set of stems:
Play all the stems together and you should hear something very close to the original.
Separation is very good now, but it is not perfect. These cases leave the most bleed, meaning a trace of one part in another:
Listen to each stem on headphones. A little bleed is often hidden once you mix the stems again.
The splitting runs in the cloud, so there is nothing to install. On our test, a 4 minute 44 second song became six stems in 56 seconds and cost 5 credits.
Personal practice and study are fine with any song. Publishing a remix or sample of someone else's music needs permission from the owners.
It is taking a finished song and splitting it back into parts, like vocals, drums and bass, so you can hear or use each one alone.
No. It is very good on clean studio mixes. Dense songs, heavy reverb and live recordings leave some bleed between stems.
Commonly four or six. Our Every Instrument Splitter can find up to about 50 kinds of instruments.
Most of our tools cost 1 credit per minute. New accounts get 5 free credits with Google sign-in. Packs start at $7.49 for 100 credits.
MP3 at 320 kbps, one file per stem.