Privacy Preserving Radio Frequency Speech Sensing With Deep Learning Towards Improved Hearing Aids
Michaela Reay, Hira Hameed, Muhammad Ali Imran, Qammer H. Abbasi · 2024
Hearing loss is a profound public health issue typically addressed with microphone sensing hearing aids calibrated to compensate for an individual's hearing loss pattern. However, microphones often generate low sound quality in noisy environments resulting in low device adoption. This paper explores use of noise resistant Radio Frequency (RF) radar speech sensing micro-Doppler (µD) shifts generated from a speaker's vocal tract with Deep Learning (DL) speech classification for audio replay. It extends earlier work using a vowels dataset to whole sentences to train and test seven DL model types that gave encouraging accuracy results ranging from 75\% to 91\%. Model training time typically took 4 minutes and no more than 20 epochs to converge with loss rates suggesting potential viability of RF µD sensing for more complex vocabularies.