Token-Level Ensemble Distillation for Grapheme-to-Phoneme Conversion
Accepted by Interspeech2019.
The code is modified from fairseq, added with knowledge distillation criterion. You can use it by setting the criterion
--criterion word_knowledge_distillation
You can check the file fairseq.md for more information.
@article{sun2019token,
title={Token-Level Ensemble Distillation for Grapheme-to-Phoneme Conversion},
author={Sun, Hao and Tan, Xu and Gan, Jun-Wei and Liu, Hongzhi and Zhao, Sheng and Qin, Tao and Liu, Tie-Yan},
journal={arXiv preprint arXiv:1904.03446},
year={2019}
}