Restoration of Face Mask-Induced Speech Intelligibility Degradation Via Neural Bandwidth Extension

초록

Since coronavirus disease 2019 (COVID-19) has prevailed in some countries, it is still mandatory to wear face masks to prevent from spreading the viruses. However, the impact of wearing face masks creates distractions and interruptions in voice communication in terms of sound quality on both perceptual quality and especially intelligibility. To provide a solution to this problem, we employ deep learning-based bandwidth extension (BWE) technique with deep neural network to restore intelligibility. Furthermore, we conduct acoustic measurement of impulse responses of various face masks to analyze the degradation effect of wearing face mask on speech sounds. The experimental results have proven effective, with a noticeable enhancement in speech intelligibility compared to mask-induced noisy speech.

키워드