I'm actually building this exact thing right now. Using Gemini TTS with its multi-speaker capabilities. Soundtrack comes from a Suno Playlist which the tool will download the entire playlist at the start.
replyPlays 3-5 songs prior to a "break", where the logic will call out channel donators vs pick a random topic to discuss, then use an LLM to generate the script before sending to gemini.
In the end; its really just a wrapper around FFMPEG broadcasting to an SRT or RTMP stream. I can upload the latest branch when I get home to share
If you intersperse the music and fill in with articles from HN or your favorite subreddits it would be a fun listen.
replythat’s great to hear. I’d love to check out your project and have a chat if you’re up for it. maybe i can contribute to it instead of working on a separate project.
replyThat sounds like a dystopia.
replyAs bearish as I am on AI, outside of the agent deciding it only wants to play "Semi Charmed Life" and nothing else, I fail to see how this is a nightmarish hellscape in and of itself.
replyI feel standard commercial radio is already that hellscape. I mostly listen to rock (classic or alternative radio stations) and it's the same ~20 things being rinsed and repeated every hour and every day. Has been this way for decades.
replyI am pretty sure 90% of playlists you hear when out and about are generated by spotify algorithms at this point. I would more be concerned because it seems like a weird use for LLM when recommendation engines are probably better than it? I suppose it has a ton of training data around music forums and people recommending songs, but it is hard to believe that would beat out billions of hours of tracked listening spotify has access too.
reply