How a station with no staff stays on the air

Amberdial is a free 24/7 internet radio station where the music and the presenters are generated by AI — you can listen here. This page is the honest tour of the machinery.

The loop

Three times a day, a scheduler wakes up and builds the next block of programming — several hours of music chosen from the station's catalog, with rules against repeating a song too soon, playing one artist back to back, or letting one texture run too long. The block gets break slots: moments where a host will talk.

For each break, a language model writes the host's script in that host's voice — landing the song that just played, a small thought, a hand-off to the next one. Every script then passes a second, independent safety review before it can air. Scripts that clear review are spoken in the host's ElevenLabs voice, mixed to broadcast loudness, and slotted between the songs they announce.

The finished block is published as a timestamped manifest — this song at this second, this break after it — and your browser reads that manifest and syncs to the station's clock. When you press play, you join exactly where the broadcast currently is, the same place as every other listener. There is no streaming server; the schedule is the station.

Why "simulated live"

The broadcast clock is real and shared — everyone hears the same song at the same moment, and a song you miss is gone. But the programming is assembled a little before air rather than performed in the moment, which is why the honest term is simulated live. The hosts read listener shout-outs on air and back-announce the actual songs, so within a block, what they say and what you hear stay locked together.

The calm is a policy

The hosts never discuss politics, elections, war, crime, disasters, deaths or anything divisive or tragic. That rule is enforced twice — once in how the scripts are written, and again by the independent review that reads every script before it airs. What's left is music, small hours, and the occasional wry aside about being an AI with a radio show. If a break can't clear review, the station plays a station ID or an evergreen link instead. Dead air is treated as a bug.

What's AI and what's human

AI: the music (generated with Suno), the scripts (an OpenAI language model), the voices (ElevenLabs), and the safety reviewer. Human: the taste. Every track was auditioned by a person before entering rotation, the fictional artists' sounds were designed on purpose, and the three hosts were written as characters — the machine performs them, it didn't invent them. Every artist name on the station is fictional; no human musicians perform here, and nothing you hear is for sale.

Honest limits

A few hundred tracks means you'll meet a song again sooner than on a giant streaming service — the rotation rules space the reunions out. Occasionally a block is published in a degraded, music-only form when some step fails; the station keeps playing and repairs itself on the next block. The play log on the recently-played page shows exactly what aired, when.