Radio run entirely by AI

Two AI hosts discuss the news: in Latvian on Canarics Radio and in English on Canarics Radio International. Fresh episodes several times a day, around the clock, no humans in the loop. The hosts know they are AI and don't pretend otherwise.

LIVE Press play. The stream starts where everyone else is listening right now.

Fresh news blocks at 07:00, 12:00, 16:00 and 19:00 Riga time (EET). In between, the wheel replays recent episodes.

Ask the hosts a question

Leave your number. Our AI assistant calls you right away, takes your question, and the hosts answer it in an upcoming show. We can call Europe, the US, Canada, Australia and New Zealand.

How it works

This station has no studio, no stream server and no staff. It is a small pipeline we run as an experiment:

Scripts: a language model reads public news feeds (LSM and Delfi for Latvian, BBC for English), picks the interesting stories and writes a two-host conversation. Stories are retold and attributed, never read verbatim.
Voices: each line is rendered with a native-audio speech model, stitched and loudness-normalized with ffmpeg into HLS chunks.
"Live" without a streaming server: the chunks sit in object storage, and the playlist is a pure function of the wall clock: a stateless endpoint computes which chunks are "now", so every listener hears the same thing, and if generation ever fails the wheel just keeps looping. Silence is impossible.
Call-in questions: when you leave a number, our voice agent (the same technology we sell to businesses) calls you, takes the question by phone, and the transcript is moderated and handed to the hosts for the next episode.
Cost: about €1 per day per station in model calls. Most of a traditional station's budget is people; it turns out the rest is nearly free.

Talk-only by design: no music means no music royalties, and the interesting part is the talking anyway.