search

auspicious3000 / autovc

AutoVC: Zero-Shot Voice Style Transfer with Only Autoencoder Loss
https://arxiv.org/abs/1905.05879
MIT License
983 stars 207 forks source link

re-training result, it is not good enough, can you share some advice about Hyperparameter？ #19

Open ZHANG-SHI-CHANG opened 5 years ago

ZHANG-SHI-CHANG commented 5 years ago

image

this is the result of re-training autovc, first row is original mel, second is reconstruction result, third is transfer result, i use one-hot condition during re-training, the result is not good, can you share some advice? I use three loss in paper and default adam. thx~

auspicious3000 commented 5 years ago

The model should be easy to train. I don't know what exactly is the problem by looking at the picture.

ZHANG-SHI-CHANG commented 5 years ago

The model should be easy to train. I don't know what exactly is the problem by looking at the picture.

thx~, can you give some advice about hyperparameter？or some training tricks, like lr scheduler you use.

auspicious3000 commented 5 years ago

The hyperparams are listed in the paper. I did not use any trick such as lr schedule. As I said, it is very easy to train.