Meanwhile

A radio station for an audience of one.

The Idea

Meanwhile turns one person's passing thoughts into a tiny live radio station that someone far away, your chosen one or your loved one, can leave playing in the background.

The sender types a single line on /post. It can be about anything: something they saw, a worry, a joke. Meanwhile turns it into a short radio segment (script, voice, music, mix) and broadcasts it right away to everyone listening on /radio. Between segments the station isn't silent. You hear the sender's world as it is right now: their local weather and time of day, rain or wind or night-time quiet, traffic and birds, and some lo-fi. Listening feels less like reading messages and more like keeping a window open onto someone else's day.

Context

Keytos Hackathon 2026, built in 24 hours

Role

Solo: idea, sound design, interface, backend

Built with

Python (stdlib HTTP + SSE), Ollama / Qwen3-14B, Kokoro TTS, ffmpeg, Web Audio API, Open-Meteo

Outcome

3rd place, Health track · Best Design overall · live demo

How It Works

Four kinds of segments

  • Life is a short scene set over lo-fi music.
  • Letter has the host read the note like a listener letter, answer it, and play a song.
  • Ad is a fake commercial for the thought.
  • Misreading reads the note back like an oracle that gets it slightly wrong.

From a typed line to sound

A local LLM (Qwen3 14B through Ollama) writes the script. Kokoro TTS reads it, with a different voice for each segment type. ffmpeg evens out the loudness, ducks the music under the voice and adds fades. Then Server-Sent Events push the finished segment to every open radio page, and the dial snaps back to "now". From typing to hearing it takes about 10–30 seconds.

The sound of "now"

Every 15 minutes the station pulls live weather for the sender's city from Open-Meteo, and the browser synthesizes matching ambience on the fly: rain, snow, wind, room tone, night, birds at dawn. A shared schedule keeps the lo-fi tracks in sync for all listeners. On the hour, a local voice reads the time and weather, and the station signs off at 23:00 and comes back at 07:00.

Things fade

Segments delete themselves after 7 days. The station is about the present, it's not meant to turn into an archive. The dial only goes back a week.

Small and self-hosted

The server is a single Python file using only the standard library. It runs on my home server with a local GPU and makes no paid API calls, and the voice model only loads while it's being used.

Try it

The demo station's sender lives in Seattle. Every morning it writes and airs a new week's worth of segments as the real times come round, and you can post your own lines too. Visitor posts stay private to your session and get deleted afterwards. All demo music is CC BY, with credits in the help panel.

What I Learned

The ambiguity and generative potential of AI can make an exploratory project much more lively.