Skip to content

Stem separation

Audionaut can split a mixed recording into its instruments: select a clip, choose Edit → Separate Stems…, and four new tracks appear — Drums, Bass, Other and Vocals — each carrying that part of the clip, lined up with the original.

Separation is powered by Demucs (Meta AI Research's htdemucs model) running on your own computer through demucs.cpp. Nothing is uploaded anywhere.

The model

The first time you separate stems, Audionaut asks to download the model — a one-time download of about 80 MB. It is kept in a Models folder in your user application-data directory (Settings → Separation shows the exact location, with a Show Folder button), and verified against a checksum before it is used.

Settings → Separation also lets you download the model ahead of time, remove it again, choose whether the original track is muted after a separation, and pick how many threads the separator may use. Each thread works on its own stretch of the clip, so more threads finish sooner at the cost of more memory; the default is the number of processor cores.

If your computer is offline, copy the file ggml-model-htdemucs-4s-f16.bin into the Models folder by hand — Audionaut checks it the same way.

Separating a clip

  1. Stop playback and select exactly one clip. The clip's track must be mono or stereo; clips up to ten minutes long can be separated.
  2. Choose Edit → Separate Stems….
  3. A progress window shows the render, the separation and the writing of the stems. Cancel stops the job and leaves the project untouched — it can take a few seconds to react.

Separation runs on the processor and takes a while: expect roughly the length of the clip per thread on a recent desktop machine, longer on laptops. The application is busy for the duration.

When it finishes, four tracks are added at the bottom of the track list, named after the clip — Take 3 - Drums, Take 3 - Bass and so on. Each holds one clip — mono for a mono source, stereo for a stereo one — starting where the source clip starts (a fade-in that reaches ahead of the clip is included, so the stems start that much earlier). The original track is muted, so what plays afterwards is the separation; unmute it to compare, or turn this off in Settings → Separation (Mute the original track after separating).

The whole operation is a single edit: Cmd+Z removes all four tracks again. The stem files stay in the project's Media/Audio folder until the project is saved.

The stems are rendered at 44.1 kHz, the model's rate, whatever the rate of the source material; Audionaut resamples them on playback like any other file.

From an agent

Agents talking to Audionaut over MCP have the same command as the separate_stems tool: it takes the project and a track (and optionally a clip), and reports the four tracks it created. It needs the model to have been downloaded from the app once.