| Title |
Unsupervised Speech Recognition via Utterance-wise Pseudo-labeling |
| Authors |
임재민(Jaemin Lim) ; 김기연(Kiyeon Kim) ; 조성현(Sunghyun Cho) ; 이석복(Suk-Bok Lee) |
| DOI |
https://doi.org/10.5573/ieie.2024.61.11.150 |
| Keywords |
Speech recognition; Unsupervised learning; Adversarial training; Pseudo-labeling |
| Abstract |
A recent study finds that the use of an auxiliary self-supervised objective significantly improves the speech recognition performance. However, this requires pre-computation of pseudo labels on the dataset, leading to training inefficiency. In this paper, we propose a new objective using utterance-wise pseudo-labeling without reliance on the training dataset. By replacing the existing objective with ours, we conduct unsupervised training for speech recognition. Our evaluations confirm the improvement not only on training efficiency but also on performance compared to prior work. |