CantusManualBack to the site

Separating stems

Back to the manual · Leer en español

Cantus can split a recording into its parts — drums, bass, vocals and the rest, or with the 6-stem model guitar and piano too — with a model that runs on your computer. Nothing is uploaded anywhere.

Separate a clip, a track or a selection

Choose 4 stems or 6 stems, Fast or Best (four passes averaged: a little cleaner, four times slower) and, on a Mac, Use the graphics processor when it can (on Windows and Linux the panel says it is not available with the runtime Cantus downloads: the model runs on the processor). Press Separate. The panel shows the progress; Cancel stops it (nothing is left behind), Hide closes the panel and the separation goes on.

How good is it? On the test songs researchers use (MUSDB18-HQ), the 4-stem model scores 8.80 dB SDR on average (drums 10.05, bass 9.78, other 6.42, vocals 8.93) — the numbers its makers published, measured again in Cantus. Busy electronic mixes are the hardest.

When it is done:

The new files go in the project's Audio/Stems folder. Their clips wear a small AI badge: each file remembers which model made it, its version and licence, when, and from which file (saved in the project).

A file at any sample rate works: it is converted to the model's 44.1 kHz and the parts back to the file's own rate and length.

The models (Settings ▸ Models)

The first time, Cantus shows the model's size, its licence and the inference runtime's, who made it and what it was trained on. Accept the licence and download fetches them:

ModelPartsDownloadLicence
HT-Demucs — 4 stems (Meta)drums, bass, other, vocals84 MB + the runtimeMIT
HT-Demucs — 6 stems (Meta)and guitar, piano (its piano is weak, its makers say)55 MB + the runtimeMIT

The runtime is Microsoft's ONNX Runtime (MIT; 15 MB on Windows, 11 MB on Linux, 37 MB on Apple silicon), downloaded once for every model, its telemetry turned off. Cantus does not host the weights: it downloads Meta's own file, checks every byte, and rebuilds the model from it. A download that is cut resumes where it stopped.

Remove deletes a model (and the runtime with the last one); it waits while a separation runs. Before a separation starts, Cantus checks there is room on the disk for it (Best needs room for four passes) and says so if not. A recording longer than one WAV file can hold as stems (4 GB: about 45 minutes at 192 kHz) is refused with a word. Opening another project cancels a separation under way.

How long it takes

About a third of the song's length on a recent laptop's processor (a 3-minute song in about 65 seconds), four times that with Best. The model runs in a helper process of its own, cantus-ml: if it fails, Cantus says so and goes on.

For AI agents

An agent connected to Cantus can call separate_stems (a track or clips; 4 or 6 parts; fast or best). It asks you first like any edit, and the model must already be installed — its licence is yours to accept.