Replicating 120h spanish dataset results

[CJai-K]

Hi @carlfm01,

Thank you for your activity in the various LPCNet/Tacotron-2 discussions. I have been trying to integrate the two models with the steps outlined by @MlWoo and yourself, but the results are not great. I can pick out words but the voice is very hoarse/noisy.

In my latest experiment I try replicating your results with the 120h spanish dataset but the results are still noisy. One things to note in my process is I used the entire dataset in training both models without making any adjustments for multiple speakers. Was this correct to do?

Another question I have is whether I need to do any special preprocessing for this dataset?

Below are my latest alignment and synthesis samples. Thank you and I look forward to your response!

samples.zip