Towards On-device Domain Adaptation for Noise-Robust Keyword Spotting

The accuracy of a keyword spotting model deployed on embedded devices often degrades when the system is exposed to real environments with significant noise. In this paper, we explore a methodology for tailoring a model to on-site noises through on-device domain adaptation, while accounting for the edge computing-associated costs. We show that accuracy improvements by up to 18 % can be obtained by specialising on difficult, previously unseen noise types, on embedded devices with a power budget in the Watt range, with a storage requirement of 1.1 GB. We also demonstrate an accuracy improvement of 1.43% on an ultra-low power platform consuming few-10mW, requiring only 1.47 MB of memory for the adaptation stage, at a one-time energy cost of 5.81 J.