An ultra-low-power always-on voice activity detector (VAD) is the key enabler of acoustic sensing in wearables. The VAD listens to the environment and wakes up the main system only when there is a right activity detected. Since most human-centric applications have infrequent activities, the VAD dominates the system power. The traditional VAD using the digital feature extractor and classifier [1] requires full-bandwidth and high-resolution data conversion before digital-signal processing, drawing a substantial power (>20μW) . Recently, the analog feature extractor shows more promises in power reduction. In [2], the analog-filter bank brings the feature-extraction power down to 1μW (Fig. 22.5.1, upper). Yet, the analog-filter bank does not support reprogramming and has a large area (∼0.1 mm ² /channel) that limits the number of input channels of the following deep neural network (DNN). The mixer-based analog filter in [4] succeeds in squeezing the feature-extraction power (142nW), but the time-interleaved operation prolongs the decision latency (512ms), and limits the extractable features (only the diagonal information on a spectrogram). In [5], the SNR-based VAD avoids the analog-filter bank, but the involved active circuitry raises the power budget and limits the performance in term of decision latency and classification rate.