Analog Acoustic Feature Extraction and Delta-Sigma Modulation based Neural Network Classification for Voice Activity Detection

This paper presents the circuit implementation of a novel voice activity detector (VAD) that performs acoustic feature extraction in the analog domain and neural network classification in the mixed-signal and digital domains. The acoustic feature extractor (AFE) has a 12-channel Mel-filterbank, which computes the energy of the raw audio signal in different frequency bands during each time frame. The classifier uses these energy values as input features and produces a ‘speech’ or ‘non-speech’ decision. The first stage of the classifier is a convolutional neural network (CNN) implemented within a first-order delta-sigma modulator, thus reducing the computational load for the digital backend. We designed all the analog and mixed-signal circuit blocks up to the schematic level in 65-nm CMOS technology, while the digital circuit blocks as Verilog-A functional modules.