- 2 stems: vocals and accompaniment.
- 4 stems: vocals, drums, bass and other.
- 5 stems: vocals, drums, bass, piano and other.
Three different ways to divide a mix
A stem is an estimated group of sounds, not a recovered recording-session channel. A song might contain many vocal takes, several guitars and multiple drum microphones. Stemix estimates broad groups from the finished recording; it cannot discover the names or exact routing of those original channels.
| Model | Estimated outputs | A useful starting point for |
|---|---|---|
| 2 stems | Vocals · Accompaniment | A vocal reference or a practice backing track |
| 4 stems | Vocals · Drums · Bass · Other | Listening to rhythm parts and rebalancing a band |
| 5 stems | Vocals · Drums · Bass · Piano · Other | Exploring a recording with a prominent piano part |
“Other” is an intentional catch-all
Guitar, strings, pads and instruments without their own output commonly share Other. This is not a separate guitar model. In the 4-stem model, piano is also part of that remainder. In a dense arrangement, an instrument can leak across several estimates. Judge the audible result rather than assuming every sound follows its label exactly.
Choose from your next action
If you will mute a singer, start with 2 stems and compare the accompaniment against the original. If you will practise bass, start with 4 and lower the bass fader. Choose 5 when isolating a piano estimate would change what you can learn from the recording. A synthesizer or processed keyboard may not behave like an acoustic piano in the model.
For a difficult passage, you can compare two model choices. Keep the source excerpt and your listening volume consistent. Listen for how much of the desired part remains, and whether removing it damages the instruments you want to keep. Avoid judging only by how quiet the isolated unwanted track becomes.
Each choice has a resource cost
Only the selected model is downloaded. The first-use dialog shows its current download requirement, so use that number rather than an old article or an estimate from another version. Cached shared components can make a later download smaller.
More output tracks also mean more decoded sample data. A 5-stem result can require substantially more memory than a single stereo file of the same length. If your first test fails on a small device, try a short excerpt and 2 stems before changing browsers or downloading additional models.
Stop when it is useful
The best choice is the one that helps your task with acceptable artifacts and resource use. For practising a part, a musical backing track can matter more than a perfectly silent mute. Keep an untouched original so you can always compare the musical context.
This guide describes the current product controls and implementation. Separation results depend on the recording, device and model; examples are not quality guarantees.
Further reading: Spleeter · Web Audio API · How the tools work