Wrong kernel sizes in the JDC model

[Rifur13]

It looks like the JDC model is implemented incorrectly. The comments don't reflect the shapes of the actual tensors, and the input to the first LSTM layer is of shape (b * 8, 31, 512) which I assum is wrong.

The wrong kernel size is used in the ResBlock layer. It's (1, 2), but should be (1, 4).
The original implementation has a kernel of (1, 4).

It looks like a bug, please help me understand if it is.

[hyzhan]

Maybe the audio is too short.