Exploring the Effect of Kernel Depth in Compact Keyword Spotting Models

Voice communication is a growing technology for human interaction with smart devices. Keyword spotting (KWS) plays a crucial role in voice communication systems because it is an always-on part that should accurately detect specific keywords to trigger and turn on other parts. In this paper, we investigate how the classification accuracy, memory footprint, and latency of a compact binary KWS architecture vary by changing the kernel channel size (depth) of its convolutional layers. The investigated architecture is a quantized fully integer architecture with 8-bit input and output data. Based on our evaluations for kernel depth between 1 to 10, the increase of depth quadratically increases memory footprint but the accuracy is saturated for depths greater than 4.