A New Training Pipeline for an Improved Neural Transducer
This work addresses speech recognition accuracy for tasks like Switchboard, but it is incremental as it builds on existing transducer methods with optimizations.
The authors tackled the problem of improving neural transducer training by comparing full marginalization with a maximum approximation, which simplified and sped up training, and found that their final transducer model outperformed their attention model on Switchboard 300h by over 6% relative WER.
The RNN transducer is a promising end-to-end model candidate. We compare the original training criterion with the full marginalization over all alignments, to the commonly used maximum approximation, which simplifies, improves and speeds up our training. We also generalize from the original neural network model and study more powerful models, made possible due to the maximum approximation. We further generalize the output label topology to cover RNN-T, RNA and CTC. We perform several studies among all these aspects, including a study on the effect of external alignments. We find that the transducer model generalizes much better on longer sequences than the attention model. Our final transducer model outperforms our attention model on Switchboard 300h by over 6% relative WER.