← Home

ChatStation

A radio station that composes itself. You type what you want to hear, a language model writes the piece, and a small timeline engine plays it with real recorded instruments into a live stream.

Ask for “slow rainy-day lo-fi” or “Charles Mingus meets Trent Reznor” and Google Gemini turns it into a structured piece: grooves, melodies, sections, key and tempo. While it plays you can change it. Add a saxophone, rate the rhythm or harmony on a pad, turn the Feel dials, or let it evolve on its own timer. An audience can join with a code, rate the music and suggest what comes next, and the host decides what gets through.

Source on GitHub →

What it does

Screens

Play: type a request and it takes over at the next bar.
Feel: change how it’s played without touching the notes.
Guests: rate an element on the pad and suggest what comes next.

Illustrative mockups of the interface, not live captures.

Run it yourself

  1. Install the audio stack on Linux: SuperCollider with SuperDirt, JACK, DarkIce and Icecast, plus Python 3 and ffmpeg.
  2. Clone the repo.
    git clone https://github.com/bobsabayesian/chat-station.git
    cd chat-station
  3. Copy .env.example to .env and add a Gemini API key. Copy darkice.cfg.example to darkice.cfg and set your Icecast password.
  4. Download the free sample libraries and build the sound banks.
    python3 tools/fetch_samples.py vsco
    python3 tools/build_samples.py
  5. Start everything.
    ./start-station.sh
  6. Open localhost:8080/admin, turn the sound on, and type what you want to hear.

The recordings are not in the repo. They come from VSCO 2 CE (CC0) and the University of Iowa Musical Instrument Samples, and the fetch script downloads them. Compositions use the Gemini API, which is a paid request with an hourly cap. A Cloudflare tunnel for letting guests join from outside your network is optional.

–Bob