Question about approaching Teacher - Student.

[phamkhactu]

Hi @erogol,

In my situation, I have a good base weight model (I call it a teacher), I have another speaker with 10 minutes audio data. Now I want to use Teacher - Student method to train new TTS model by one shot learning. I've used voice conversion, but it doesn't keep style speaker (another speaker) I think that my approach is:

1. Load again pretrained weight(Teacher)
2. Freeze text encoder: freeze_encoder = True
3. Train again new speaker with 10 minutes data
   How do you think about my ideas ? I need add another freeze block, does it? I would like to hear another solutions?

Thanks in advance!