The Representation Of Speech Variability And Variation In Deep Neural Networks