Deep Neural Networks for Automatic Speaker Recognition Do Not Learn Supra-Segmental Temporal Features