A comprehensive study and benchmark of 4 different Prediction Networks in RNN-T for offline

Two methods allow almost full recovery of the performance loss due to data deidentification