Single-channel speech enhancement using improved progressive deep neural network and masking-based harmonic regeneration.