Low-Power Low-Cost Audio Front-End for Keyword Spotting

Abstract

This paper presents a low power audio front end for keyword spotting. A multi-stage approach is used to reduce the power consumption of the system by only using different stages when they are required. A working prototype was created and tested to verify its functionality. The effectiveness of the multistage approach is shown by comparing the power consumption of the system in its idle state to the systems active state. The prototype has a power consumption of 4.1 mW in the idle state that can be reduced below 3 mW with a keyword detection accuracy of 87 %.