This paper presents an ultra-low-power voice activity detector (VAD). It uses analog signal processing for acoustic feature extraction (AFE) directly on the microphone output, approximate event-driven analog-to-digital conversion (ED-ADC), and digital deep neural network (DNN) for speech/non-speech classification. New circuits, including the low-noise amplifier, bandpass filter, and full-wave rectifier contribute to the more than 9x normalized power/channel reduction in the feature extraction front-end compared to the best prior art. The digital DNN is a three-hidden-layer binarized multilayer perceptron (MLP) with a 2-neuron output layer and a 48-neuron input layer that receives parallel event streams from the ED-ADCs. To obtain the DNN weights via off-line training, a customized front-end model written in python is constructed to accelerate feature generation in software emulation, and the model parameters are extracted from Spectre simulations. The chip, fabricated in 0.18-mu m CMOS, has a core area of 1.66 x 1.52 mm(2) and consumes 1 mu W. The classification measurements using the 1-hour 10-dB signal-to-noise ratio audio with restaurant background noise show a mean speech/non-speech hit rate of 84.4%/85.4% with a 1.88%/4.65% 1-sigma variation across ten dies that are all loaded with the same weights.