Vocoder training

[sankar-mukherjee]

Thank you for sharing this nice work.

I have a question.

Do i have to train the hifigan from scratch or i can use may be pretrained hifigan from nvidia ngc ? If this true then do i have to change audio processing parameters to match hifigan?

[shivammehta25]

Just add ckpt_path 'python matcha_tts/train.py ckpt_path='

[shivammehta25]

I've used the pretrained one from HiFiGAN repo, if its any different with audio parameters then I will suggest changing the parameters to match the vocoder's parameters.

[sankar-mukherjee]

Thank you for clarification on both issues. Another question, can i finetune a pretrained model with small amount of data with this repo? Where should i put the pretrained model path?

[shivammehta25]

Just change the filelist location in dataset and resume training with ckpt_path

[shivammehta25]

I am closing this for now, feel free to reopen incase of further questions.