Lightweight Streaming Keyword Spotting System Based on Resonate-and-Fire Neurons

Deep Neural Networks (DNNs) have significantly advanced keyword spotting (KWS) accuracy. However, deploying those KWS models on edge devices requires both substantial memory footprints and significant economic costs. To address this, we propose a lightweight streaming KWS system that exploits the energy-efficient characteristics of Spiking Neural Networks (SNNs). Our system incorporates three innovative components. First, we introduce a spike encoding method based on Resonate-and-Fire (RF) neurons, which provides an energy-efficient alternative to the Fourier Transform for auditory time-frequency analysis. Second, we propose a frame-level streaming KWS model that processes a single frame sequence per timestep and incorporates an early stopping strategy to improve decision-making efficiency. Finally, we quantize the streaming KWS model to lower bit-widths, further reducing computational energy consumption. Extensive experiments on the Google Speech Commands (GSC) Dataset (Version 1 and Version 2) demonstrate that our approach not only maintains competitive performance compared to other SNN-based KWS models but also achieves this with fewer parameters and a reduced memory footprint. The efficiency and efficacy of the proposed system highlight its potential as a promising avenue for edge devices.