In this work we propose two postprocessing approaches applying convolutional neural networks (CNNs) either in the time domain or the cepstral domain to enhance the coded speech without any modification of the codecs. The time domain approach follows an end-to-end fashion, while the cepstral domain approach uses analysis-synthesis with cepstral d…
keras cnn generative-adversarial-network gan post-processing mfcc speech-processing wasserstein-gan speech-enhancement speech-reconstruction 1d-convolution
-
Updated
Mar 8, 2020 - Python