Does VAE model supports Tacotron 2 ?

[rishikksh20]

Nice implementation of VAE tacotron paper though I havn't read paper yet ,but start model training with LJSpeech. As I see this code is extending the Tacotron 1 implementation so I am bit curious that weather same VAE implementation possible with Tacotron 2 ?

Note : On paper, author used Tacotron 2.

[yanggeng1995]

yes, the experiments in this paper were carried out based on tacotron2, and I also implemented vae on tacotron2. Due to some research relations, it cannot be made public for the time being. You can refer to tacotron1 for migration

[rishikksh20]

Ok thanks, just a thing, the hyperparameter of VAE remain same for Tacotron 2 as of this repo or some thing needed to tweak for better performance? And roughly how much step/iteration it will take to converge with current implementation (for LJSpeech 22.5k) ?

Thanks in Advance :+1:

[yanggeng1995]

My experiment with tacotron2 just got going, and it hasn't worked out yet, so you can listen to the example the author gave(http://home.ustc.edu.cn/~zyj008/ICASSP2019/)

[rishikksh20]

util and datasets folders are not uploaded, I have copied from keithito's implementation. But still get :

    from util.util import shape_list, vae_weight 
ImportError: No module named 'util.util'

may be some file not uploaded.

[yanggeng1995]

Sorry, I forgot to upload these two folders. I just uploaded them。