A Sub-400-nW Real-Time Event-Driven Spectrogram Extraction Unit in 28-nm FD-SOI CMOS for Keyword Spotting Application

Abstract

Considering the power-hungry nature of speech processing, a keyword spotting (KWS) unit, used to detect multiple spoken words, is often integrated as a front-end layer. KWS systems are always active, and thus, it is extremely important to optimize the devoted power budget. In this context, this article presents a programmable low-power event-driven real-time spectrogram extraction unit tested for the KWS application. This chip, fabricated in 28-nm FD-SOI CMOS technology, has been combined with a software-defined convolutional neural network to demonstrate the recognition of 11 audio classes (ten keywords + background + unknown) with an accuracy equal to 87.9% and an activity-dependent power consumption measured at 391.6 nW, for a 12-keyword/min average speech rate.