Show HN: Llm.fm – an AI-generated, satirical radio show in the browser | Hacker News Reader

Show HN: Llm.fm – an AI-generated, satirical radio show in the browser

llm.fm

8 pointsryandotelliott4 comments

I've spent the past few weeks building *LLM.FM*, a GTA-inspired Radio Show where every topic, song, and advertisement is generated entirely by AI.

I loved the absurd style of the radio stations in the Grand Theft Auto games, and I wanted to see how close I could get with todays AI stack. While it isn't quite as vulgar, I think it's still a unique experience!

How it works

* GPT-4.1 writes a "high-level" show overview given a desired structure, this is done to keep coherence between segments

* Individual segment transcripts are generated with the provided outline

* ElevenLabs turns the into voices and returns word-level timestamps for a live transcript

* Segments are converted to HLS chunks, then all chunks are merged with audio padding; Song are pre-generated Suno songs based on ChatGPT generated song prompts

* Next.js + Vercel for the web app; Railway with workers to handle new show generation

The show is fully free - new episodes release at 8 AM, 12 PM, and 4 PM ET daily.

Open articleView on HN

This is crazy good. Both coherent, and consistently pretty believable, and the two songs in the episode I listened to were actually nice to listen to.

Getting the songs good, or the coherent script, or the TTS quality individually would already be good, but getting the full package so seamlessly consistent is amazing.

So far I'd say the only very minor weakness is that some of the narration is too emotionally flat. For the host, that mostly works and fits the nature of the kind of soothing radio host voice you'd expect from this kind of show (but slightly more excitement, or change in tone when changing subject would be great to prevent it from just becoming background I tune out), but it doesn't quite hit for the ads which are often over-the-top excited on radio.

But that's an incredibly minor nitpick. I've heard radio programs far worse than this on actual radio stations.

Please, please do a more extensive writeup of this (or release the code, though I fully understand if you intend to keep this to yourself). Heck, release the writeup as an episode :)

I could really see this turned into something to let users design a profile for a custom show, blending in a summary of news and media specific users care about.

@dang, if you see this, only 8 points for this post in 20 hours is criminal - if anything deserves a round in the second chance pool this is up there...

Thanks for the feedback! Thought the post was dead, so it’s great to see someone found it and left such a detailed review!

I’m going to try and open-source it in the next few weeks once I return from a trip.

Do you have anything you’d like to see in a write up?

I'd try to post it against after a reasonable time - or at least when you have a writeup. The writeup might well get the best traction.

I think what's most interesting to me is how you ensure the consistency. The "cascade" of generating a high-level plan, and then drilling down is something I've experiented with ad-hoc, but seeing it come together so nicely with the combination of multiple tools is fantastic, and so concrete examples of prompts going from the high level, to the songs, ads, etc. would be great in a writeup.

Damn it. I listened to the whole thing..

Well done. Very much enjoyed the interface

Reply on news.ycombinator.com