Why the norm consistency loss can help to preserve the speech/silence intervals of generated samples and thus decrease the noises?

[yl4579]

Because silence has a lower norm (i.e. volume) than speech part, and hence the model learns to distinguish between noise and speech by introducing this loss term.