DEV Community

Vijay Kumar
Vijay Kumar

Posted on Edited on

The Listener

#devchallenge #hf26challenge

The Listener: a ghost story that listens back

Ghost in the Machine submission. Open-weight model + open TTS + open speech recognition, all on one laptop, no internet required.

I built a ghost story that is written live from what your room does while you sit in it. Headphones on, lights off. A local model (Llama 3.2 via Ollama) writes a creepypasta one beat at a time, as a tired narrator posting updates to a night-owl forum. Your microphone is the other author. A knock on the wall becomes "something bumped the other side of the wall, softly, like it was checking the wall was still there." A long silence becomes "the refrigerator stopped humming." Say "is anyone there" out loud, and a few beats later the hallway answers in your voice, slightly wrong. Look away from the screen, and when you look back the door is open wider. Piper narrates it, and the whole thing is rendered to a video you can download.

Github:

GitHub logo Vijaykr35 / The-Listener

THE LISTENER

A ghost story that listens to you. Sit in the dark with headphones on. The story is written live, one beat at a time, by an open-weight model running on your own machine, and it is built from what your room actually does while you wait: a knock, a creak, a long silence, the words you say out loud, the moment you look away from the screen. It is narrated aloud and rendered as a video you can download. When it ends you get the whole thing as a creepypasta forum thread, plus a case file of what the microphone really heard.

Run it

./run.sh            # installs deps, fetches the voice, pulls a model if Ollama is present, starts the app
# then open http://127.0.0.1:8000  (Chrome or Edge recommended)
Enter fullscreen mode Exit fullscreen mode

Or by hand:

pip install -r requirements.txt
python scripts/get_voice.py          # Piper narrator voice (one-time download, ~60 MB)
ollama
…
Enter fullscreen mode Exit fullscreen mode
View on GitHub

The turn. The story plants a psychological reversal at the start. My favourite: The sounds are yours. The narrator puts the timestamps of the night's strange noises into a column and realizes none of them happened in their apartment. Those are your real timestamps, from your real room. At the end, the app hands you the forum thread and a case file of what the microphone actually heard.

Why open wins here.

  • Privacy. An always-on mic is as private as input gets. With an open stack, the audio never leaves the machine. There is no API key and no account.
  • Offline. Unplug the network. It still works.
  • Free. Zero cost per story.
  • Control. A 3B model will happily wander off and explain the monster. So the model doesn't run the story: a director holds a hidden storyboard, an act structure and a dread curve, and a rule checker verifies every beat in code (never name the thing, one wrong detail at a time, planted details must pay off, the rule is broken exactly once, end unresolved). Bad drafts go back to the model with the exact violations; after two misses, a template writer built from the same bible steps in. I could only build that loop because I could see and swap every piece. 600 simulated stories pass every rule.

Stack: Ollama (Llama 3.2 3B), Piper TTS, faster-whisper, FastAPI, vanilla JS + canvas + MediaRecorder.

Top comments (0)