Speech Enhancement for Demodulated Signals under Multipath Fading Communication Channels

In analog communication channels, such as radio broadcasting, the superposition of multiple reflected signals causes multipath fading. Multipath fading often results in the fluctuation of the received electric intensity levels of these signals; thus, it causes severe quality degradation in audible sounds. In this paper, we focus on speech enhancement under a fading communication channel with additive Gaussian noise. We attempt to reconstruct the original speech based on the use of denoising autoencoders that employ mean-squared-error and additive perceptual evaluation of speech quality (PESQ)-based loss functions in multi-task learning (MTL). The experimental results indicate that the MTL-based autoencoder improves PESQ scores from 2.00 to 2.75 for demodulated signals under fading communication channels with additive Gaussian noise.