Self-initiated project at Cerevo Inc, Oct. 2022

The Secret Box is a box which opens when the secret knock pattern was detected, just like we played spy game in childhood.

I saw once someone in the internet built an Arduino-based door opener. The device is attached to a door and listens door knock pattern, turns the door handle when the "right" nock pattern was detected. I thought that I could also build similar knock detector with embedded machine learning.

Recently I had opportunity to play around with the EdgeImpulse. EdgeImpulse is a extremely neat-yet-powerful service to build ML module for embedded devices using TensorFlow Lite.
I have collected almost 10 minutes of sound samples for each correct data (labelled "Knock") and wrong data (labelled "Rhythm"), as well as supplement data sets "Noise" and "Unknown" from Google Speech Commands Dataset.

Result and Current problem
I have successfully set up the ML module to detect "my secret knock pattern" in high probability. The box is probably not very practical, but funny to store some candies in it.
I also have learned a lot about "data engineering", which is quite different paradigm from conventional coding.

Material: Wio Terminal, 3D printed box, Servo motor
Link to the original blog post in Japanese.