A Low-cost keyword spotting architecture based on wavelet packets feature extraction for edge devices

Sayed Salehi and Prakash Dhungana
University of Kentucky


Abstract

This paper proposes a novel voice keyword spotting (KWS) architecture that uses wavelet packets to reduce the implementation cost of its feature extraction component. The approach achieves a 54% reduction in latency and a 32% decrease in memory compared to conventional Fourier-based KWS architectures.