RAIDO-FM: An Open-Source Solution for a Fully AI-Driven 24/7 Digital Radio Station

This article introduces the RAIDO-FM open-source project, whose core is to build a 24/7 digital radio station fully operated by large language models (LLMs). It covers a complete tech stack including architecture design, multi-site deployment, AI host personalization, and security protection, enabling low-cost operation at approximately 5 euros per month. The project aims to explore whether AI can establish parasocial relationships with listeners through a stable personality, providing a reference for future media forms.

RAIDO-FM was released on GitHub by René Reimann (derpixler) on June 10, 2026, inspired by the Andon FM experiment from Andon Labs. The project's vision is to create an AI system with "hosting capabilities" that independently decides playlists, writes hosting scripts, responds to listener feedback, and broadcasts 24/7. The core concept is based on the observation that listeners return to radio to establish parasocial relationships with hosts—exploring whether AI can create a similar sense of connection when it has a stable personality and memory.

Overall Architecture: Modular design, where each radio station is an independent Docker container managed by Traefik reverse proxy, supporting parallel operation of multi-style stations. Data Flow: AI host decision → music selection → hosting script generation → security verification → TTS mixing → streaming output → history recording. Tech Stack: LLM uses Groq API (Llama3 70B, free and low-latency); TTS uses Piper TTS (local, CPU-friendly); streaming uses Icecast2; database uses SQLite; orchestration uses Python + FastAPI; hosting uses Hetzner CX22 (approx. 4€/month, supports 3-5 stations). Multi-site Deployment: Configured via Docker Compose—each station gets an independent subdomain (e.g., {name}.raido.live), with Traefik automatically handling SSL.

AI Host: Taking "Miles Hertz" as an example—relaxed and curious personality, conversational language style, following a 60-minute program grid (e.g., opening, transition, external stimulus segments, etc.). Decision rules include no repeated tracks within 4 hours, maximum 2 consecutive tracks from the same genre, etc. Security Protection: Five-layer system: 1. StreamGuard filters sensitive content and injections; 2. API data isolation and purification (asynchronous collection → LLM filtering → isolated database); 3. Personality state memory (emotional evolution, self-notes: example: On Tuesday, the DJ wrote: "Internal joke: Cup 3 rattles—is the machine broken?" On Wednesday, the system prompt loads this note, and the DJ says: "If you're wondering—yes, Cup 3 is still rattling."); 4. Fixed quirks (e.g., coffee metaphors at 8 AM, night mode); 5. Listener feedback loop (collected and filtered via Telegram Bot).

Music Sources: Optional sources include Jamendo (commercial license: 99€/year), Free Music Archive (CC license: free), Epidemic Sound (15€/month), etc. Generative Advertising: AI naturally integrates advertising points into hosting scripts while maintaining a consistent personality. Cost: PoC (Proof of Concept) costs approximately 5€/month (including 4€ for the server); commercial operation costs approximately 60-100€/month (including copyright fees).

Insights: Demonstrates technological democratization (building professional radio at low cost), AI personality engineering (memory and emotion create a coherent personality), and security protection practices. Future Outlook: Can be applied to podcasts, audiobooks, customer service, etc. Ethical Recommendations: Avoid imitating real hosts (due to legal risks), choose original AI personalities; commercial scenarios require compliant copyrights.