Meatball can see, hear, and talk back
A caseless home server built from other people's cast-offs: salvaged GPUs, an abandoned-office pile of audio dongles, a '60s microphone, early-'90s speakers. Taught to see, hear, think and speak. Every model runs on the metal and nothing touches the cloud.
What it can see right now
It can just ask
Meatball says hi through the Altec Lansings, then listens on the always-on mics. A reply → probably a person. Silence → probably not. It also does this on its own when the cameras catch someone, the simplest sensor there is.
It tracks its own motion
each camera, every ~10s · left = frame + heatmap (where it changed) · right = the raw subtraction · line = motion over time
Always-on ears
no transcriptions yet. Speak near a mic and it'll trip the floor and land here
Four cast-off parts, four working senses
Eyes
Two scrounged webcams. One grabs a frame a minute and serves it live; the vision model narrates what is in front of it.
Ears
A '60s garage mic, a thrift-store condenser, and the webcam mics, feeding a local Whisper that transcribes the room.
Mouth
Early-'90s Altec Lansings driven by a local neural voice. First word lands in about a third of a second.
Brain
A local LLM on two old GPUs, plus a self-calibrating audio rig. No keys, no cloud, ever.
The bill of materials
Nothing here was bought new for this.
| The junk | What it became |
|---|---|
| No case at all | open-air on the bench, a body you can reach into |
| Two old GPUs | run the local LLM, Whisper STT, and a neural voice, on the metal |
| 20 assorted second-hand external drives | the memory |
| A pile of Plantronics ADACs from an abandoned office | the audio lanes: speaker and mics over USB |
| An old factory start/stop button | the power switch |
| Altec Lansing speakers, early '90s | the mouth |
| A mic from my grandfather's garage, '60s | an ear |
| A Realistic condenser mic, Salvation Army | another ear |
| A baseless Salvation Army monitor | the display |
| Scrounged Logitech webcams | the eyes, and it turned out the best ears too |
Written up as it was built
- I gave a junk-pile eyes, ears, and a voicefield note 01. The whole saga: the salvaged bill of materials, the debugging traps that cost an hour each, the calibration insight, and finally talking to the WOPR out loud.
- The math of listeningfield note 02. Low-level DSP from raw samples to a working voice gate: the real FFT, windowing (200x less leakage), spectral-subtraction denoise and its U-curve. Every number from a live run.
- Teaching the eyes to ignore a box fanfield note 03. One cheap fan fooled both senses. Locking the camera's auto-exposure, then an adaptive per-cell gate that self-mutes the fan, the monitors and lighting blips, so it only fires on what should not move.