search

descriptinc / melgan-neurips

GAN-based Mel-Spectrogram Inversion Network for Text-to-Speech Synthesis
MIT License
980 stars 214 forks source link

Melgan multispaeker pretrained model fine-tuning with single speaker #39

Open MuruganR96 opened 2 years ago

MuruganR96 commented 2 years ago

Melgan pre-trained model fine-tuning with a single speaker (720 samples - 1 hour audio samples)

it didn't work. the loss keeps increasing after the 30th epoch, and generated samples are not good

May I know any reason for this? and How can we overcome this issue?

Kindly request, Please give your suggestions.

Thanks