GuangChen2016

@Durham Hello, Durham. Are you still working on this? I still couldn't got good results, and the alignment seemed didn't work according to the alignment pictures. Do you have some...

Sorry, miss that

@MaxMax2016 你好，我这边是先训了个多人的平均模型，然后在小数据集上进行adapt，adapt后的效果还可以。但是在adapt后的模型的基础上用了teacher的wave encoder&Flow，并且冻结他们网络参数来训练student模型，得到的模型音质下降的很明显。请问这个训练过程有什么问题么？另外这里训student的时候判别器的参数是随机初始化还是用的teacher的。

@MaxMax2016 加上这个loss后，有比较高add_blank和不加add_blank以及是不是用sdp duration的差异么

@feng-yufei Hi, I met the same issues using my datasets, what's your conclusions now?

@feng-yufei Yeah, and did you use loss_kl_r now and set the coefficient to 0.01 for your training or just remove loss_kl_r?

@OnceJune 可以听下效果么