search

CODEJIN / Glow_TTS

An implement of GlowTTS model. Several modes are added: speaker embedding, prosody encoder(GST), and gradient reversal.
MIT License
51 stars 12 forks source link

The mismatch of sampling rate of Glow-TTS acoustic model and the vocoder #9

Open actuy opened 1 year ago

actuy commented 1 year ago

Hi, I found that the sampling rate when training the Glow_TTS model is set to 24000 (in pretrained checkpoint), while the pretrained checkpoint in the given vocoder repo (https://github.com/CODEJIN/PWGAN_Torch) set the sampling rate to 16000.

This mismatch leads to bad audio outputs. Would you please share the pretrained vocoder with the same sampling rate?

Thanks a lot in advance!

CODEJIN commented 1 year ago

Dear @actuy,

Hello. I apologize for the delayed response. After the issue was raised, I searched the correct vocoder and checkpoint. Although I couldn't find a version updated to 24K directly in PWGAN repository, I found that the following repository was set up: https://github.com/CODEJIN/GAN_based_Vocoders.

It's been a while, so I don't have a clear memory, but it seems that there was some confusion during the process of integrating the original PWGAN project into the new repository. The PWGAN 24K model is included in the repository, and the checkpoints are also uploaded at Google Drive.

If the model doesn't produce the desired output even with these checkpoints, I'm truly sorry, but currently, I don't have a way to recover it. I appreciate your understanding.

I hope this model can be helpful.

Best regards,

Heejo