A 28-nm MFCC-Free Keyword Switchable Keyword Spotting (KWS) System With Transferred Training Algorithm

Abstract

In this brief, we propose an ultra-low-power mel frequency cepstral coefficients (MFCCs)-free keyword switchable KWS system that supports ten sub-classifiers (2 keywords each, 20 keywords in total) through a time-domain transferred training convolutional neural network (TT-CNN). The proposed TT-CNN reduces the model size by sharing the first two convolutional layers with all the keywords with a transferred training approach. Hence, the power budget for memory and computation is largely reduced. The TT-CNN supports flexible keyword demand in different scenes by selecting different kernels in the custom-designed 5T-SRAM. The time-domain feature of the proposed TT-CNN avoids the power-hungry feature extractor (FEx), further reducing the overall power consumption. To benchmark with the state-of-the-art, we demonstrated the proposed system with two cascaded scalable 10-Class KWS chips in 28nm CMOS. Our design achieves a high accuracy of 92.8% on 20 keywords from the Google speech command dataset (GSCD). It also shows that the memory overhead for each keyword can be reduced by 20% with the lowest reported 20-class KWS power consumption of $1.2~\mu $ W.