A 28.5µW All-Analog Voice-Activity Detector

We describe the design principles behind an all-analog, neural-network-based voice activity detector (VAD). Our architecture, which uses a neural engine with weights that only take on values -1,0 and 1, achieves high accuracy along with low latency. Implemented in a 130 nm CMOS process, the VAD consumes 28.5 μW from a 1.8 V supply and achieves a latency of only 10 ms. The measured hit rate for speech(non-speech) is 94.3%(94.1%).