@inproceedings{chi2021align,
 abstract = {Non-autoregressive encoder-decoder models greatly improve decoding speed over autoregressive models, at the expense of generation quality. To mitigate this, iterative decoding models repeatedly infill or refine the proposal of a non-autoregressive model. However, editing at the level of output sequences limits model flexibility. We instead propose iterative realignment, which by refining latent alignments allows more flexible edits in fewer steps. Our model, AlignRefine, is an end-to-end Transformer which iteratively realigns connectionist temporal classification (CTC) alignments. On the WSJ dataset, Align-Refine matches an autoregressive baseline with a 14x decoding speedup; on LibriSpeech, we reach an LM-free testother WER of 9.0% (19% relative improvement on comparable work) in three iterations. We release our code at https://github.com/amazon-research/align-refine.},
 address = {Online},
 author = {Chi, Ethan A and Salazar, Julian and Kirchhoff, Katrin},
 booktitle = {Proceedings of the 2021 Conference of the North {A}merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)},
 month = {jun},
 publisher = {Association for Computational Linguistics},
 title = {Align-Refine: Non-Autoregressive Speech Recognition via Iterative Realignment},
 url = {https://nlp.stanford.edu/pubs/chi2021align.pdf},
 year = {2021}
}