Skip to main content

Before you install

The station brings almost everything with it, but it plays your music rather than its own, and it runs in Docker. This page is what to have ready, so neither install page stops halfway to send you off for something.

What you need​

  • Docker on a linux/amd64 machine. Either Unraid with Community Applications, or any Linux host with Docker and the Compose plugin (with Docker Compose). Docker Desktop on a Mac or a PC works for trying it out.
  • Somewhere the music comes from. These tutorials use a Spotify account, which has to be Premium because the station fetches the audio, or a media server of your own: Navidrome or another Subsonic server, Jellyfin, Emby, or Plex. Have the server's address to hand, and an account on it (Navidrome), an API key and a user name (Jellyfin and Emby) or your Plex token (Plex). YouTube Music, Pandora and Apple Music work too, as the library describes. The station holds no music of its own.
  • Internet access on the first boot, for the image itself and for the speech model's weights. The weights are downloaded the first time the station speaks, not when the image is pulled.

What you can add later​

  • A language model. Without one the presenter talks from the station's own phrasings, so a station with none still talks. With one, the model writes what the presenter says. It can be Anthropic, Gemini, or anything that speaks the OpenAI API, Ollama on your own network included. Give the station a model covers it.
  • A mail server, for sign-in links and emailed codes. Without one, you sign in with the password you set.

How big a machine​

These figures are from a fresh full container, idle, before any library was synced:

full
Memoryabout 2 GB, of which the speech server is about 1.4 GB
The image on diskabout 9.5 GB
The data directoryabout 400 MB, most of it the speech model's weights

Give it 4 GB of memory, so the station has room to work when it renders speech and measures records. The data directory then grows with what the station does: records are cached before they air, and the cache grows to the size of the library you play. Install § Two kinds of thing explains how to put that cache on a different disk.

latest is the same size without the database's share. slim brings no speech server, so it saves its memory and most of its disk. The voice then runs on another machine, which is where a graphics card helps. The station itself needs none.

On a Mac with Apple Silicon​

The image is built for amd64 only. Docker Desktop runs it under emulation, which works and is slower, most noticeably the first boot and rendering speech. It is fine for trying the station out. For a station that stays on the air, use an amd64 machine.

Next​

Install on Unraid, or install with Docker Compose, where the short way is one command.