noise vector in MelGAN and learning an accurate conditioning

[acids-ircam]

Hello,

There is something confusing in the paper, although it's in overall greatly written.

2.1 Generator architecture, it states that additional noise is not fed in the generator since it does not affect the perceptual quality of the result

2.3 Training objective, eq (1) and (2) state G(s,z) meaning that Generator processes both spectrogram and gaussian noise vector

If you inject noise, where is that happening please ?

This brings me to an other question. I read GAN-TTS, they inject noise when conditioning hidden activations wrt. speaker embedding. And they use conditional discriminator to ensure that the audio is both realistic and accurately conditioned.

If there is neither noise in MelGAN, nor conditional discriminator, how do you assess that the generator is learning and generalizing for the conditional generation please ?