Should z_q_dur drawn from Gaussian distribution?

[trinh-hoang-hiep]

I have a concern about whether the random variable z_q_dur should follow a normal distribution or not. When assuming z_q_dur follows a normal distribution, of course z_audio will follow a normal distribution, but with normal flow when z_q_dur is nonGauss, z_audio can still be Gauss. So maybe the first KL function should be a little more flexible?