Stems

Take a mixed recording apart into its instruments. Select a clip, run one command, and six new clips appear underneath it: drums, bass, vocals, other, guitar and piano. Separation runs entirely on your machine.

What it does

Stems: Separate Selection runs the selected clip through a six-stem BS-RoFormer model and puts each stem back on the canvas as a real audio clip, grouped under the source and in playing order — drums, bass, vocals, other, guitar, piano. The source clip is muted rather than removed, so you can unmute it to compare against the original at any time.

The whole result is one edit: six clips, the group and the mute all go in together, and a single ⌘Z puts you back where you started.

The clip is separated dry — as recorded, before its effect chain. This is deliberate. Separating the processed signal would bake the clip's reverb into every stem, and then playing those stems back through the same chain would apply it a second time.

Using it

1. Select one audio clip.
2. Run Stems: Separate Selection from the palette or from a chat.

One clip per run. If you have several selected, it names the one it is acting on and leaves the rest alone.

There is no window. The chat you ran it from is where it reports progress, which on a long clip is the more useful place for it to be.

What it needs

• Your project, for writing — it inserts the six stem clips.
• Your audio, for reading — it renders the selected clip to separate it.
• One network download. On first use it fetches the separation model (336 MB) from huggingface.co, once. The download goes through the host, against the host's allowlist, with the SHA-256 carried in the signed catalogue. Every run after that is offline.

Separation is not fast and it is not free — expect real time on a long clip, and watch the chat for progress.

Platforms

macOS, Apple Silicon only. The model runs on a prebuilt ONNX Runtime that has no Intel Mac build.