Pretrained models about wavernn

[fatchord]

I'll be uploading a pretrained model today or tomorrow.

[geneing]

I have a pretrained model using LJSpeech dataset in my fork of this repo: (https://github.com/geneing/WaveRNN).

[fatchord]

@loretoparisi I haven't trained the models on any open source datasets so I'm not sure what the legal situation is with the weights I have regarding distribution. I'll err on the side of caution for now - sorry.

[loretoparisi]

@fatchord thank you, this is a good point. Having a domain transformation from the input dataset to the model, this will depend on the data usage agreement you have and any point about data transformation that could or could not be mentioned. By the way, assuming you have followed the paper, which was the dataset structure? This could help me to find open source datasets and do a training from scratch. Thank you very much.

[rishikksh20]

@fatchord how much hours of audio needed to generate good voice like your demo voice for TTS (with Tacotron 2)?

[fatchord]

@rishikksh20 I used around 20 hours of data. One thing to note is that I used Tacotron 1 for those samples in my demo voice (only difference is that I didn't use Griffin-Lim at the end).