Train/test split used for VCTK data

[spun-oliver]

Thank you for making your code and pretrained models available.

I would like to use your pretrained VCTK models for benchmarking in an evaluation. I assume that the VCTK models at https://drive.google.com/drive/folders/1-eEYTB5Av9jNql0WGBlRoi-WH2J7bp5Y are the ones used for the work presented in Section 4.3 of the HIFI-GAN paper (https://arxiv.org/pdf/2010.05646). The paper mentions that nine speakers were randomly held out from training: it would be helpful to know which speakers these were. The 5 samples given in the "Unseen Speakers (VCTK Dataset)" section of https://jik876.github.io/hifi-gan-demo/ are from from VCTK speakers p226, p271, p226 (again), p318 and p292; and also the samples seem to be from the mic1 version of the data, which I assume was used for training. It would be great if you could provide the IDs of the other 5 speakers, or the training.txt/validation.txt files that were used to train these models.

Thanks!