Ultra High Frequency/ Very High Frequency (UHF/VHF), additive stationary, additive non-stationary, codec-based distortion
RATS dataset
Log Spectral Distance (LSD) and Multi-Scale Spectral Loss (MSSL)
Speech2Speech-Augment
PatchNCE loss
Aggregating noise with clean speech followed by codec distortion using g726
5.78
Magnitude and Phase. The GANs act on the magnitude component
Content Masks, Foreground Attention Masks and Background Attention Masks
The SpeechAttentionGAN is the least practical to use for general application as it requires parallel training data and getting parallel data may be difficult in real world applications
The original converter was MelGAN. it had to be changed because upon conversion of a noisy spectrogram to an audio clip, MelGAN performs poorly and outputs unintelligible clips
