Using a vision-inspired keyword spotting framework, we propose an architecture with input-
dependent dynamic depth capable of processing streaming audio. Specifically, we extend a …