A 291nW Real-Time Event-Driven Spectrogram Extraction unit in 28nm FD-SOI CMOS for Keyword Spotting Application

Considering the power-hungry nature of speech processing, a keyword spotting (KWS) unit, used to detect single spoken words, is often integrated as a front-end layer. KWS systems are always active, thus it is an extreme importance to optimize the devoted power budget. In this context, this paper presents a programmable low-power event-driven real-time spectrogram extraction unit tested for KWS application. This systemon-chip (SoC), fabricated in 28nm FD-SOI CMOS technology, has been combined with a software defined convolutional neural network (CNN) to demonstrate the recognition of 11 audio classes (10 keywords + background noise) with an accuracy equal to 90.08% and an activity-dependent power consumption measured at 291. 6nW, for a 12 keywords/minute average speech rate.