Sub-mW Keyword Spotting on an MCU: Analog Binary Feature Extraction and Binary Neural Networks

Keyword spotting (KWS) is a crucial function enabling the interaction with many ubiquitous smart devices in our surroundings, either activating them through wake-word or directly as human-computer interface. For applications, KWS entry point for interactions device and, thus, an always-on workload. Many are mobile and their battery lifetime heavily impacted by continuously running services. similar services thus focus when optimizing overall power consumption. This work addresses energy-efficiency on low-cost microcontroller units (MCUs). We combine analog binary feature extraction neural networks. By replacing digital preprocessing proposed front-end, we show that energy required data acquisition can be reduced $29\times $ , cutting its share from dominating 85% to mere 16% of consumption reference application. Experimental evaluations Speech Commands Dataset system outperforms state-of-the-art accuracy efficiency, respectively, 1% notation="LaTeX">$4.3\times 10-class dataset while providing compelling accuracy-energy trade-off including 2% drop notation="LaTeX">$71\times reduction.