Exploiting the compressed spectral loss for the learning of the DEMUCS speech enhancement network