Created by: alanbekker

Let's assume my audio segment consists of 8000 samples at 16khz (0.5sec) In the preprocessing stage using a sliding window of 20msec and stride of 10msec I receive 50 different spectrograms..you can reach this number by yourself (500msec/10msec=50)

When I run prediction I can see that the softmax output consist of 17 distribution vectors.

Why 17? I assume this is a result of context window width of +-1 (and then 50/3=~17), but I have no reference that the context width indeed is +-1..can someone correct point out a reference for this or correct me?

Thanks in advance!