A 31-Feature, 80nW, 0.53mm2 Audio Analog Feature Extractor based on Time-Mode Analog Filterbank Interpolation and Time-Mode Analog Rectification

Abstract

To alleviate the feature extraction bottleneck in always-on, on-device Keyword Spotting (KWS), we propose two novel analog circuit techniques that are combined into an efficient analog feature extraction architecture: 1) Time-Mode Analog Filterbank Interpolation (TM-AFI) uses digital XOR gates to double the number of outputs of an analog filterbank, 2) Time-Mode Analog Rectification (TM-AR) uses a single digital XOR gate as an analog full-wave rectifier. Among other analog feature extractor chips using a software classifier for a KWS demo, the 31-feature, 80nW, 0.53mm2 prototype is 18× more power-efficient and 3.3× more area-efficient than the most area- and power-efficient published works, respectively, while maintaining competitive >90% accuracy on 10 keywords.