A Keyword-Spotting(KWS) Chip Featuring a Bio-Inspired Neuron Model in 65-nm CMOS

In this study, we present a Keyword Spotting (KWS) system using supervised Spiking Neural Networks (SNNs). We specifically focused on the Leaky Integrate-And-Fire (LIF) neuron model and found that a model incorporating a convolutional spiking layer achieved competitive performance with conventional Deep Neural Networks (DNNs), attaining a 94.63% accuracy for the 4-class Google Speech Commands Dataset (GSCD). Moreover, by leveraging the sparse spiking activity of the LIF neuron model and combining it with a Fast-Sampling Convolutional Neural Network for feature extraction, we have significantly reduced power consumption of the chip to 7.7µW while maintaining accuracy. The achieved latency is 18.74 mS/decision, with a core area of 0.917 mm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>, including the complete KWS model.