Robots that hear through touch. Tap, scratch, or shake something near your microphone — a mug, your desk, a pen, a pill bottle — and teach this two contact sounds by example. A tiny model runs on your device, turns each tap into a timbre fingerprint, and learns to tell them apart. Then it listens and names what you touch. Sound reveals material and hollowness that a camera can't — this is that, in your own hands.
What the microphone hears · live spectrum
Name each sound, hit Record a tap, and tap the object once. Add a few examples each — 3–4 is plenty. Then just tap: the fingerprints below are how it tells metal's bright ring from wood's dull thud.
A nearest-prototype classifier over a 26-D timbre fingerprint (24 log-frequency bands + brightness + ring-out time), Node-verified to separate two contact timbres at 100% from 5 examples each. Honest few-shot, fully in-browser — the transparent cousin of what Duke SonicSense and Stanford ManiWAV learn from contact audio. · Institute for Physical AI @ BMI