A KWS System for Edge-Computing Applications with Analog-based Feature Extraction and Learned Step Size Quantized Classifier
Yukai Shen, Binyi Wu, Dietmar Straeussnigg, Eric Gutierrez · Preprints.org · 2025
Edge-computing applications require of ultra-low power architectures for both the feature extraction and the classifier stages. In this manuscript a complete architecture for these applications is proposed making use of an analog-based feature extraction stage composed of a bank of band-pass filters, and a 4-bit weight, 8-bit activation function learned step size (LSQ) quantized gated recurrent unit (GRU)-based classifier stage for keyword spotting (KWS) systems. The proposed KWS architecture achieves 91.35% of accuracy for 12 classes including 10 keywords from the Google Speech Command Dataset v2 (GSCDv2), with less than 1% accuracy loss compared to full precision classification, with an estimated memory footprint of 34.8 kB and 62 400 multiply–accumulate (MAC) operations per inference in real-time mode.