A 16nm 140TOPS/W 5μJ/inference Keyword Spotting Engine Based on 1D-BCNN

• Published in: IEEE transactions on circuits and systems. II, Express briefs Express briefs, 2023-06, p.1-1
• Main Authors: Lin, Tay-Jyi, Ting, Yi-Hsuan, Hsu, Meng-Ze, Lin, Kuan-Han, Huang, Chung-Ming, Tsai, Fu-Cheng, Sheu, Shyh-Shyuan, Chang, Shih-Chieh, Yeh, Chingwei, Wang, Jinn-Shyan
• Format: Article
• Language: English
• Subjects: 1-dimensional binarized convolutional neural network; Clocks; Convolution; Convolutional neural networks; Current measurement; energy efficient; Engines; Keyword spotting; low power; Random access memory; Speech processing
• ISSN: 1549-7747
• DOI: 10.1109/TCSII.2023.3290230

This brief presents an event-driven keyword spotting (KWS) system for reducing the significant but usually ignored energy dissipations on the "always-on" A/D converter and microphone. "Low energy per inference" and "fast responsiveness" are new design goals of such KWS engine. A 7-layer 1-dimensional binarized convolutional neural network (1D-BCNN) was designed to achieve 95% inference accuracy for detecting 10 keywords, plus silence and unknown, from raw speech, and 64 32-element signed binary inner product units were allocated in the engine to deliver the 4,096 operations/cycle maximum throughput. The 16nm implementation consumes only 0.1mm2 silicon area and 5μJ/inference energy (including memory accesses), while achieving 1.72ms response time. The performance is comparable to state-of-the-art KWS designs without sacrificing number of detectable keywords or inference accuracy.