What It Is
The AI radio is two halves: 4ort.live is the station; 4ort-music is the factory feeding it. A generation daemon runs around the clock: rotated LLMs act as the songwriting department, inventing each track's theme, BPM, key, language, title, caption, and lyrics, and ACE-Step v1.5 (an 8-step SFT-merged turbo variant, running as a GGUF on CPU) synthesizes the music itself. No human picks what gets written; the factory composes continuously.
What makes it a library rather than a noise firehose is the curation: at a sustained pace of ~50 tracks a day, roughly 80% get culled. Only the keepers earn a place in the catalog, each one titled, with synced lyrics, ratings, and keep flags. The result so far: 100+ published original tracks, streaming 24/7.
It's also engineered like software, not like a jam session: every track carries a batch id that snapshots the exact generation regime that produced it (the prompt templates, the LLM rotation, the theme pools, the model configuration). Changing the regime means opening a new batch, so the library's history stays reproducible and comparable across eras of the factory.
Highlights
- 24/7 generation daemon: ACE-Step v1.5 synthesizing original music continuously on CPU, no GPU required
- LLM songwriting department: rotated models invent each song's theme, BPM, key, language, title, caption, and lyrics
- A hard quality bar: ~50 tracks/day generated, ~80% culled, so only keepers reach the catalog
- Versioned generation regimes: every track records the exact templates, model config, and theme pools that made it
- 100+ published tracks: each titled, with synced lyrics and per-user ratings and favorites
- A programmable library: stations, queue, now-playing state, and live broadcast events exposed over a REST + SSE API, agent-ready like the rest of the lab
Status
● Live and generating around the clock. Hear the library on air: