More info on training

[ajaiswal1008]

I am training the voice conversion model from scratch using the Obama audio file. I have trained for around 20K steps and the loss is not decreasing much (recon0:0.03, recon:0.03, vocoder:0.04). Also the audio file generated after 20K steps sounds like Obama's voice but the content information is lost. Can you advice on what steps should i take going forward? Should i just wait till 60K steps as mentioned in the paper. Also what are the loss values that would indicate a good model performance.

Thanks in advance