Design and Implementation of a Low Power Switched-Capacitor-Based Analog Feature Extractor for Voice Keyword Spotting

Abstract

A low-power keyword spotting (KWS) is demanding for the smart human-device interface. The conventional analog feature extractor utilizes an analog filter bank that consumes large power and area. This paper presents a KWS with a low-power switched-capacitor-based feature extractor. The feature extractor employs two pipelining time-domain convolutional neural networks (TD-CNNs) to extract sufficient features for KWS. The TD-CNNs utilize sparsity aware computation (SAC) and sparsified quantization (SQ) for a 4-bit weight quantization. The features are quantized to 2-bit for further classification by an off-chip deep neural network. The analog feature extractor is designed in a 55-nm CMOS process and post-layout simulation is provided. It consumes 4.4 µW at a 1.2-V power supply, with an area of 0.39 mm2. It achieves an accuracy of 92.2% for the google command dataset (GSCD) with five classes.