Use Tacotron2 to make voice cloning

[Lizh-TW]

Hi!
I've been trying to use Tacotron2 to do voice cloning experiments for the past few months.
I encountered some problems and would like to ask you

1. I have the problem in training at the first, because of embedded_speakers = aux_input["d_vectors"], at the test step. It loads all of the .wav files to clone, so I guess this is wrong, then I modified to embedded_speakers = aux_input["d_vectors"][:, 0, :]. Error generating test sentences on multi-speaker #760 (comment)
2. I have trained once it in the old version (v0.2.x and 100 people) successfully, although there has a difference between the two speakers, the result is not good.
   -speaker 1
   -speaker 1 clone
   -speaker 2
   -speaker 2 clone
3. But after once I trained successfully, then when I update to v0.3.x and trained a new model. the sound between the two speakers has no difference anymore, I didn’t modify important parameters, so I wonder to know is my problem related to the version or not.
4. Finally, I want to ask, what is the difference between 2 and 3 here? I'm confused, how to set my config?

• speaker1
• speaker2
• speaker3
• config.json.txt
•