A 250.3μW Versatile Sound Feature Extractor Using 1024-point FFT 64-ch LogMel Filter in 40nm CMOS

A $250.3 \mu \mathrm{~W}$ always-on sound feature extractor that facilitates general-purpose sound recognition AI processing encompassing 35 -word voice command recognition, environmental sound recognition, and musical instrument recognition is developed. Conventionally, approximated melfrequency cepstrum coefficients (MFCC) feature extractors composed of a limited number of FFT samples (256 points), and filter channels (10 channels) are utilized for energy reduction; however, their applicability is restricted to wake-up word recognition resulting in high NRE costs. To overcome these challenges, we developed a LogMel filter feature extractor employing a 1024-point FFT and 64-channel Mel filter bank, which enables versatile applications across a diverse range of sound recognition tasks including 35 -word voice command recognition. To minimize circuit area and power consumption, three techniques are employed: (a) radix-<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> single-path delay feedback ($\mathrm{R}^{2}$ SDF) which uses serial FFT processing for circuit area reduction, (b) zero-skipping Mel filter bank for a $1 / 25 x$ circuit area reduction by storing and accumulating only nonzero elements, and (c) Log LUT, an LUT approximation to reduce the number of cycles by a factor of 20 compared with the CORDIC implementation. Designed and implemented in a 40nm CMOS process, the proposed extractor demonstrates a power efficiency of $14.9 \mathrm{~nJ} / \mathbf{f r a m e} /$ word for a 35 -word voice command recognition task, showcasing a $1.1 \times$ improvement in power efficiency and a $17.5 \times$ increase in the number of recognizable voice commands compared to state-of-the-art KWS-specific simplified MFCC audio extraction circuits.