Deep residual spiking neural network for keyword spotting in low-resource settings.

We propose a practical solution for the implementation of keyword spotting (KWS) system on portable devices, that features all three properties required for battery-powered portable scenarios: low power usage, small footprint, and high accuracy. In particular, we study an end-to-end KWS system with deep residual Spiking Neural Network (SNN), perform experiments on Google Speech Commands Dataset, and compare with both state-of-the-art ANN and SNN models. First, the proposed solution outperforms its ANN counterpart and other SNN in terms of energy efficiency. Second, it requires a smaller footprint (86.5K) than other ANN and SNN (210K) models. Third, in terms of classification accuracy, it outperforms the existing power-efficient SNN benchmark by 4% to 17%. The proposed solution is an example of the unparalleled performance of spiking neural network in real-world applications.