Most AI audio tools read a script you wrote. deadair runs a station: a director builds the running order, picks the records, and writes what the DJ says between them from real facts (Wikipedia, news, charts, weather), then speaks it and streams it over Icecast. Personas have memory and stories that carry across shows. Your library, your LLM, your voice, all pluggable. Self-hosted in one container, MIT licensed, with listener apps for iOS, Android and desktop as well as a stream deck plugin.
I've been a Spotify user for a long time now. I was really excited when they introduced their AI DJ but it never really lived up to what I thought it could. I've also had this idea in my head since I was in high school but until the rise of AI, I never would have been able to build this myself.
I took the approach of "plugin first" where the core functionality is a queue based system aka the "Director". The music, enrichment, llm, voice, rss (news), weather, podcasts, etc are all plugins. This approach allows others to extend the station, for example, a youtube music provider could be built (I don't have an account to test). There's also an api if anyone is interested in extending it that way.
I've been daily driving this for a few months now so I'm fairly confident that it's "stable" but happy to take issues and ideas to improve it. I also run it on Unraid so it's been released there as well.