A Charge-Trap-Transistor-Based Fully Analog Machine Learning Inference Engine for Audio Keyword Spotting

We present an analog machine learning chip for audio keyword spotting based on charge-trap transistors in 22 nm FD-SOI process. This implementation does not require any analog-to-digital conversion and uses 4-bit weights with a full scale of 200 nA, resulting in an ultra-low-power solution suitable for edge devices. A recurrent neural network has been implemented on chip demonstrating a 96% correlation between silicon measurement and simulation, while consuming an average of 10 nJ per inference.