Analog Features Extractor for Ultra-Low Power Embedded AI Listening and Keyword Spotting

In this paper, a novel ultra-low power audio system is designed and tested on keyword spotting (KWS). It implements mixed (analog-digital) processing and allows always-on detection with an average power consumption of < 77 mu W, detecting one word every 30 seconds. Combining mixed signal processing, it achieves 88% classification accuracy between 8 classes. It is based on an always-on Analog Features Extractor (AFE), which provides spectral information, and main processor wakes-up only when necessary for minimal power consumption. Then, this information is sampled by a low frequency ADC to compose a spectrogram processed with a Convolutional Neural Network (CNN) for classification. It can be used for long term environmental monitoring or intelligent Internet of Things (IoT) using small batteries such as a single CR2032 coin cell to reach up to 1 year of autonomy.