A 5.6µW 10-Keyword End-to-End Keyword Spotting System Using Passive-Averaging SAR ADC and Sign-Exponent-Only Layer Fusion with 92.7% Accuracy

This paper presents a 10-keyword end-to-end keyword spotting (KWS) system. Passive averaging effectively improves the analog front-end (AFE) SNR with a small power overhead of 20nW. A sign-exponent-only layer fusion scheme reduces the model size and multiplication power overhead by 63.5% and 29.8%, respectively, maintaining the KWS accuracy. Fabricated in 65nm CMOS, the design provides the highest accuracy of 92.7% and the lowest power consumption of 5. 6Jl W compared to prior arts aimed at 10 keywords.