Learn about innovative domain-adaptation methodology for improving keyword spotting systems in noisy environments through this technical talk presented by ETH Zürich doctoral student Cristian Cioflan. Explore how on-device adaptation techniques can enhance accuracy by 15-25% by refining models using prerecorded speech data augmented with on-site noise samples. Discover the implementation details of this approach on the GAP9 ultra-low power platform, which achieves superior performance with just 10 kB memory adaptation cost compared to larger baseline networks. Gain insights into addressing the critical challenge of accuracy degradation in keyword spotting systems when complex environmental noises corrupt input signals.