Fine-tuning procedure for mb_melgan vocoder, Voice Quality degrading with Fine-tuning.

Hello! I’m trying to finetune a Vocoder model for Indian accents. I’ve followed the suggestions from thread #296 and have arrived at a suitable acoustics model.

To improve the output voice quality of the present vocoder( multiband_melgan.v1) model, I had followed the finetuning process mentioned in examples/multiband_melgan with 940000.h5 multiband_melgan.v1-EN as the pretrained model.

However the output has degraded(completely muffled speech) compared to the pretrained vocoder.

I had used the same dataset as the one used for fastspeech model training, with this command,

python ./examples/multiband_melgan/train_multiband_melgan.py \
--train-dir ./dump/train/ \
--dev-dir ./dump/valid/ \
--outdir ./examples/multiband_melgan/exp/train.multiband_melgan.v1/ \
--config ./examples/multiband_melgan/conf/multiband_melgan.v1.yaml \
--use-norm 1 \
--pretrained mb_melgan_generator.h5 

These are the loss plots that I obtained,

Please help me debug this problem, thank you

[WadoodAbdul]

Thanks for the clarification @dathudeptrai 😃

[dathudeptrai]

Thanks for the help, @dathudeptrai. I Had not continued after 200k steps.

Just for clarification, I have to tune the Generator+Discriminator for the rest of the steps(1M), not just the generator for 1M steps. Is that correct?

yes 😄