A 510-nW Wake-Up Keyword-Spotting Chip Using Serial-FFT-Based MFCC and Binarized Depthwise Separable CNN in 28-nm CMOS

• Published in: IEEE journal of solid-state circuits Vol. 56; no. 1; pp. 151 - 164
• Main Authors: Shan, Weiwei, Yang, Minhao, Wang, Tao, Lu, Yicheng, Cai, Hao, Zhu, Lixuan, Xu, Jiaming, Wu, Chengjun, Shi, Longxing, Yang, Jun
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.01.2021; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; Binary neural network (NN); Circuit design; CMOS; Computation; Convolution; data reuse; depthwise separable convolution (DSC); Feature extraction; Hardware; Iron; keyword spotting (KWS); Leakage; Mel frequency cepstral coefficient; near-threshold voltage (NTV) design; Neural networks; Power consumption; serial fast Fourier transform (FFT); Task analysis; Threshold voltage; Word processing; Words (language)
• ISSN: 0018-9200; 1558-173X
• DOI: 10.1109/JSSC.2020.3029097

We propose a sub-<inline-formula> <tex-math notation="LaTeX">\mu \text{W} </tex-math></inline-formula> always-ON keyword spotting (<inline-formula> <tex-math notation="LaTeX">\mu </tex-math></inline-formula>KWS) chip for audio wake-up systems. It is mainly composed of a neural network (NN) and a feature extraction (FE) circuit. For significantly reducing the memory footprint and computational load, four techniques are used to achieve ultra-low-power consumption: 1) a serial-FFT-based Mel-frequency cepstrum coefficient circuit is designed for FE, instead of the common parallel FFT. 2) A small-sized binarized depthwise separable convolutional NN (DSCNN) is designed as the classifier. 3) A framewise incremental computation technique is devised in contrast to the conventional whole-word processing. 4) Reduced computation allows a low system clock frequency, which enables near-threshold voltage operation, and low leakage memory blocks are designed to minimize the leakage power. Implemented in 28-nm CMOS technology, this <inline-formula> <tex-math notation="LaTeX">\mu </tex-math></inline-formula>KWS consumes <inline-formula> <tex-math notation="LaTeX">0.51~\mu \text{W} </tex-math></inline-formula> at a 40-kHz frequency and a 0.41-V supply, with an area of 0.23 mm 2 . Using the Google speech command data set, 97.3% accuracy is reached for a one-word KWS task and 94.6% for a two-word task.