SpideySense.
Loading the actual CADFour printed parts + reference electronics
Two INMP441 microphones on the glasses capture synchronized stereo audio.
Zipformer scores the sound. Timing differences give a rough direction.
The priority engine sends an icon and arrow to the 128 × 64 OLED.
The main body holds the Pi and optical guides. A stepped lid covers the wiring. One microphone carrier sits on each arm of the glasses.
Download the four STLsOpen full CAD & wiring explorerCAD shown here is the actual local export. Printed fit, comfort, and optical focus still need physical checks.
The same sound reaches one microphone a fraction of a second before the other.
GCC-PHAT estimates that delay. The original simulator below shows the two-mic headset swaying as it walks, with the head-yaw signal underneath.
Original renderer and Python simulation output. Public playback uses generated signal replays.
Every second, Zipformer analyzes the last two seconds of audio and scores 527 AudioSet categories. Thresholds, cooldowns, and priority rules turn those scores into alerts.
Inference runs on the laptop. The Pi streams the audio and receives the display frame. The model weights have not been fine-tuned.
418 source recordings · 1,672 total cases. These are offline, per-recording results from the repo’s benchmark, not room accuracy. Model scores are not probabilities; small sample sizes limit the result.
Full report & methodology