Questions about training from scratch

[Levent9]

Hello, I followed the steps in run.sh to train with the LibriTTS-R dataset. Below is the training process loss. When I use the current checkpoint to synthesize speech, it is almost noise. Based on the loss, does the training of the model appear normal? Thank you!

Loss: nll_loss: reg_l1_loss: reg_l2_loss:

[ZhihaoDU]

I think the training process is normal. On LibriTTS corpus, I got losses like these after about 106K iterations:

loss=6.557
nll_loss=4.271
reg_loss=2.286
reg_l1_loss=1.733
reg_l2_loss=2.840
out_acc_1=0.167
out_acc_2=0.100

If the synthesized speech is almost noise, there may be other bugs ?

[Levent9]

Thank you very much for your response. The text-to-speech synthesis is now functioning normally.