A 1.5-μW Fully-Integrated Keyword Spotting SoC in 28-nm CMOS With Skip-RNN and Fast-Settling Analog Frontend for Adaptive Frame Skipping
Heejin Yang, Ji-Hwan Seol, Rohit Rothe, Zichen Fan, Qirui Zhang, Hun-Seok Kim, David Blaauw, Dennis Sylvester
Abstract
We propose a fully integrated low-power keyword spotting (KWS) system on chip (SoC) with content-adaptive frame subsampling, implemented in 28-nm CMOS technology. The system is co-optimized from end-to-end including the analog frontend (AFE) and digital backend with a skip-recurrent neural network (RNN) KWS algorithm. The SoC performs dynamic power gating based on the decision from the skip-RNN algorithm that allows opportunistic frame skipping to reduce the power consumption without compromising the KWS accuracy. The design employs a fast-stabilizing AFE, enabling fast OFF to ON transitions with a settling time of less than 1 ms. A low-power feature extractor (FE) and RNN classifier sprint with a relatively fast clock to minimize the latency of the frame-skipping decision and to minimize the leakage power overhead. The SoC integrates a custom-designed latch-based always-on ON-chip memory to reduce leakage power to store all RNN weights on the chip. The proposed system achieves 1.48 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\mu \text{W}$ </tex-math></inline-formula> with an average of 76% skip ratio across frames, achieving 92.8% accuracy on a 7-class subset of the GSCD dataset. This work represents a significant step toward a low-power KWS SoC with content-adaptive frame subsampling for energy-efficient, deep-learning-enabled Internet-of-Things (IoT) devices.