A Low-Power Hardware Accelerator of MFCC Extraction for Keyword Spotting in 22nm FDSOI
Liyuan Guo, Matthias Jobst, Johannes Partzsch, Stefan Scholze, Andreas Dixius, Matthias A. Lohrmann, Seyed Mohammad Ali Zeinolabedin, Christian Mayr · 2023
With the development of artificial intelligence, the real-time feature extraction of acoustic signals is required in a wide variety of applications, such as keyword spotting and speech recognition. Feature extraction based on Mel-frequency cepstral coefficients (MFCCs) is one of the most significant methods thereinto. A software implementation of the MFCC extraction results in relatively high power consumption and computational time limitation, often making it unsuitable for tiny battery powered devices. Therefore, an on-chip accelerator of MFCC extraction is of interest in cutting-edge scenarios. This paper presents a fixed-point low-power hardware accelerator of MFCC feature extraction implemented in 22nm FDSOI technology. It consumes an average power of 2.78µW for 1024-sample frame at a clock frequency of 1MHz. For keyword spotting, the quantized accelerator achieves an average accuracy of around 96% working along with different classification networks.