An all-analog, neural network based Voice Activity Detector (VAD) architecture is proposed which eliminates the quantization error of analog to 1-bit digital conversion and achieves high accuracy along with low latency. This VAD is implemented in a 130nm CMOS process and it consumes 28.5 μW from a 1.8V supply. It achieves a hit rate speech/non-speech of 94.3%/94.1% and an overall latency of 10ms.