Post #3588558

By the time it gets to layer 18 (the output from which gets fed into direct signal-to-sequence translation CTC), the two models end up with a very different looking output.