Enhanced De-Essing via Neural Networks

Abstract

De-essing is the process of attenuating vocal sibilance in audio recordings. Especially in audio mastering, conventional de-essers often degrade the clarity of the source signal due to unreliable differentiation between vocal sibilance and other high-pitched sounds. Machine learning poses a promising solution to this problem. In this context, a new de-essing approach based on a convolutional neural network architecture is presented. The introduced prototype de-esser outperforms existing de-esser plugins in terms of erroneous signal attenuation and was rated favorably by audio professionals.