This paper presents a novel analog/mixed-signal classifier for voice activity detection (VAD) in 65-nm CMOS technology. We propose several new circuit blocks, such as a current copier based mixed-signal multiplier, a scalable current mode max-pooling circuit, and an analog rectified linear unit (ReLU) circuit, which we used to implement a convolutional neural network (CNN) based machine learning (ML) model. The weights of this model were computed via offline training using open source datasets. The trained VAD was tested on multiple audio files by carrying out system-level simulations in MATLAB and transistor-level simulations in Cadence. Both MATLAB and Cadence simulations produce identical outputs, and we observe a classification accuracy of 90-95%.