A Low-Power Keyword Spotting System With High-Order Passive Switched-Capacitor Bandpass Filters for Analog-MFCC Feature Extraction

This paper presents a low-power high-accuracy key-word spotting (KWS) system based on analog passive switched-capacitor (SC) bandpass filters (BPF). The proposed system inno-vatively extracts the Mel-frequency cepstrum coefficient (MFCC) features with all-analog circuits, providing a better spotting performance than the present analog short-time amplitude or energy features under the same condition. At the circuit level, the analog-MFCC extraction includes the low-noise amplifier, BPF, squarer, integrator, and discrete cosine transformer. And thanks to the effective analog-MFCC features, the size of the fully-connected neural network (FCNN) classifier in our KWS system is enormously reduced. A high-order and fully-differential BPF is also proposed, achieving ultra-low power and high dynamic range by combining zero and pole generation stages rather than building stages separately in traditional ways. Fabricated in 0.18-
$\mu \text{m}$
CMOS, our filterbank of eight BPFs is measured with a power consumption of 83.2 nW, with a 69.7 dB dynamic range at 5% THD, having advantages over other SC BPFs with similar functions. The total power consumption of our feature extractor is 661.7 nW, achieving an accuracy of 96.6% in two-keyword spotting by an FPGA-based, 15k bit parameter FCNN with a 9.6
$\mu \text{s}$
latency.