NVIDIA/waveglow

Citation of waveglow_256channels_v5

philippaichinger opened this issue · 4 comments

philippaichinger commented

Hi,
is there a recommended way of citing published model parameters? I am using waveglow_256channels_v5. Also, I would be interested in some hard facts regarding this model parameters (what training data, how long trained etc.)
Best, Philipp

rafaelvalle commented

You can cite the paper and the GitHub repo or the Mellotron paper which describes this waveglow model.

v5 was trained on a studio quality single female speaker dataset with about 20h. We've trained the model for about 1.5M iterations.

  • ❤️1
philippaichinger commented

OK thanks. Is it really 1.5 mio. iterations (within which 1 batch of a few training samples is used), or epochs (within which the network sees each training sample once, i.e., all batches are used once)?

rafaelvalle commented

1.5M iterations

anderleich commented

Hi,

How did you decide to stop the training process? Is there any metric you used or it was just inference audio quality?

  • 👍1