The RNN transducer loss is a prototype feature, see `here <https://pytorch.org/audio>`_ to learn more about the nomenclature. It is only available within the nightlies, and also needs to be imported explicitly using: :code:`from torchaudio.prototype.transducer import rnnt_loss, RNNTLoss`.