TY - GEN
T1 - Sound Event Detection on A Single MCU
T2 - 3rd International Conference on Sensing, Measurement and Data Analytics in the Era of Artificial Intelligence, ICSMD 2022
AU - Chen, Haolin
AU - Zhu, Maiteng
AU - Zhou, Jingtao
AU - Qiao, Jiaqing
AU - Liu, Bing
N1 - Publisher Copyright:
© 2022 IEEE.
PY - 2022
Y1 - 2022
N2 - Sound event detection (SED) in Internet of things (IoT) domain always face the challenges of strict power consumption constraints. To achieve ultra-low power, many SED systems use complex analog circuit to extract sound feature and construct algorithm, which discourages the portability and generality of the algorithm to be deployed on other IoT devices, and also takes up more spaces. In this paper, we propose a light-weight SED algorithm based on long-term speech diversion (LTSD) feature extraction and support vector machine(SVM) classifier to realize SED fully on a single Microcontroller Unit (MCU) without depending on any extra analog frontend. The algorithm achieves accuracy rates of 86.3%, better than other lightweight SED algorithms. Onboard deployment is realized on a single MCU CC1350, achieving 88.2% recognition accuracy within 3.125 μs speed, which only occupies 33.4KB flash and 9.94kB SRAM, and consumes 23.1mW power, reflecting its ability to be portable on other tightly memory-constrained and low-power IoT devices.
AB - Sound event detection (SED) in Internet of things (IoT) domain always face the challenges of strict power consumption constraints. To achieve ultra-low power, many SED systems use complex analog circuit to extract sound feature and construct algorithm, which discourages the portability and generality of the algorithm to be deployed on other IoT devices, and also takes up more spaces. In this paper, we propose a light-weight SED algorithm based on long-term speech diversion (LTSD) feature extraction and support vector machine(SVM) classifier to realize SED fully on a single Microcontroller Unit (MCU) without depending on any extra analog frontend. The algorithm achieves accuracy rates of 86.3%, better than other lightweight SED algorithms. Onboard deployment is realized on a single MCU CC1350, achieving 88.2% recognition accuracy within 3.125 μs speed, which only occupies 33.4KB flash and 9.94kB SRAM, and consumes 23.1mW power, reflecting its ability to be portable on other tightly memory-constrained and low-power IoT devices.
KW - Internet of Things
KW - Long-Term Speech Diversion
KW - Sound Event Detection
KW - Support Vector Machine
UR - https://www.scopus.com/pages/publications/85150451974
U2 - 10.1109/ICSMD57530.2022.10058217
DO - 10.1109/ICSMD57530.2022.10058217
M3 - 会议稿件
AN - SCOPUS:85150451974
T3 - 2022 International Conference on Sensing, Measurement and Data Analytics in the Era of Artificial Intelligence, ICSMD 2022 - Proceedings
BT - 2022 International Conference on Sensing, Measurement and Data Analytics in the Era of Artificial Intelligence, ICSMD 2022 - Proceedings
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 22 December 2022 through 24 December 2022
ER -