About model implementation differences

[sourcesur]

Hi, thanks for your effort and sharing the code!
The architecture blocks in the speech prompted text encoder and CFM decoder differ from the initial ones introduced in the paper. I would like to know what made you do the changes. Was the model not converging with official architecture?

[p0p4k]

Just easier to implement and more modular. Keeping it open ended to make it more accessible to do experiments.

[sourcesur]

I wanted to reproduce the results from the paper, so I used this repo (master branch) to train the model on LibriTTS. I trained it for 800k steps and longer but the overall generation quality is quite far from the official demo.
Have you tried reproducing the results?

[p0p4k]

try to play around with the speech prompt encoder. I have not trained this model yet.