Does the fast step can generate 1s in about 0.8s as the paper said?

[azraelkuan]

Avg time per step inference on CPU: 0.002576863145828247
Avg time per step inference on GPU: 0.003681471061706543

0.003681471061706543 * 16000 = 59s

so how can we get the result in the paper?

[erogol]

I don't know. I asked the authors via an email but got no reponse. Maybe you could reimplement it with dilated convolution instead of two convolutions for each part of the input. But right now, I have other things todo over FFTNet and ll return to it shortly.

[azraelkuan]

@erogol yes, i have finish the basic implement. Thanks!

[erogol]

@azraelkuan let me know if you have any update, question or comment. Free to discuss :)

[DLZML001]

Hi, any update on this issue?

[erogol]

@DLZML001 Nope, I went to WaveRNN now.