Choice and effect of "segment size"

[BenjSta]

My question relates to the "segment size" parameter. You use 8192, which is 372 ms @ 22050 Hz sampling rate. If I have computed correctly, the receptive field width is roughly 300 ms (?) in the v1/v2 configurations. That would mean that during training most of the generated audio is affected by padding in the convolutions. How did you choose the "segment size" parameter? Is there a trade-off between reducing the effect of padding and achieving enough speaker variability within each batch in multi-speaker training. Or does the padding even act as a regularization?

Regards