DeepClean receptive field

The receptive field of each sample in the output of the standard DeepClean convolutional autoencoder architecture is 96 samples on either side of it. For a sample rate of 4096Hz, this amounts to a window of ~47ms over which DeepClean is capable of conditioning its predictions.

While I think there are good reasons to prefer architectures that can encode entirely local information, this feels much too local and is likely to limit both what frequency ranges we're capable of modelling (e.g. a 20Hz signal can't even make a full cycle in this amount of time) and the bandwidth of signals we can model (I'm sure the uncertainty principle is at work somewhere here, I'm just not clever enough to express exactly what that limit is).

It might be worth exploring deeper DeepClean architectures, or ones that pursue more aggressive striding or pooling to increase this receptive field.

No file, test, or entry point is named. Start by locating the standard DeepClean convolutional autoencoder architecture and measuring its receptive field; done would require agreeing on and evaluating a deeper, more aggressively strided or pooled architecture.