yl4579 / StyleTTS2

StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models
MIT License
4.38k stars 340 forks source link

asr negative loss #236

Open yijingshihenxiule opened 1 month ago

yijingshihenxiule commented 1 month ago

Hi, @yl4579 ,thank you for your awesome work. I met some problems when train my model. When I train asr model with my phoneme symbols, I got negative ctc loss and negative train loss. Is it normal? Could you give me some suggestions?