This work explores the effectiveness of detecting positions of out-of-vocabulary words (OOVs) in a decoded utterance using attention weights and CTC per-frame outputs of an end-to-end system predicting word sequences. We show that the end-to-end approach can be effective for the task of OOV detection. CTC alignments are shown to provide better temporal information about the positions of OOV words than attention, and therefore are more suitable for the task. The detected positions of OOV occurrences are utilized for the recurrent OOV recovery task in which probabilistic representations of the pronunciations of the detected OOVs are clustered in order to find repeating words. Improved detection results are shown to correlate with better performance of the recovery of recurrent OOVs.
Cite as: Egorova, E., Vydana, H.K., Burget, L., Černocký, J. (2021) Out-of-Vocabulary Words Detection with Attention and CTC Alignments in an End-to-End ASR System. Proc. Interspeech 2021, 2901-2905, doi: 10.21437/Interspeech.2021-1756
@inproceedings{egorova21_interspeech,
author={Ekaterina Egorova and Hari Krishna Vydana and Lukáš Burget and Jan Černocký},
title={{Out-of-Vocabulary Words Detection with Attention and CTC Alignments in an End-to-End ASR System}},
year=2021,
booktitle={Proc. Interspeech 2021},
pages={2901--2905},
doi={10.21437/Interspeech.2021-1756}
}