A Low-Power Keyword Spotting Chip with Multiplier-Free MFCC Feature Extractor

Abstract

Mel-frequency cepstral coefficients (MFCC), an FFT-based speech feature extraction (FEx) algorithm, is a significant power consumer in low-power keyword spotting (KWS) chips. This work presents a KWS chip with an energy-efficient FEx, with an expanded-3bit-twiddle FFT (E3bT-FFT) algorithm which reduces power of FFT by 5.7x. Meanwhile, a multiplier-free MFCC (MF-MFCC) is proposed, effectively eliminating power-hungry multipliers and reducing the MFCC computational load by 7.3x. Fabricated in a 65-nm CMOS process, the chip occupies 0.17 mm² and consumes 2.3 µW, with the computation unit in FEx consuming just 76 nW, and achieves 94.9% accuracy on a 1-Word KWS with Google Speech Commands dataset (GSCD).