There have been some changes to CTC loss and the general style of implementation for speech workloads in TensorFlow since the Fathom release.