TY - GEN

T1 - On scaling contrastive representations for low-resource speech recognition

AU - Borgholt, Lasse

AU - Tax, Tycho M.S.

AU - Havtorn, Jakob D.

AU - Maaløe, Lars

AU - Igel, Christian

N1 - Publisher Copyright: © 2021 IEEE.

PY - 2021

Y1 - 2021

N2 - Recent advances in self-supervised learning through contrastive training have shown that it is possible to learn a competitive speech recognition system with as little as 10 minutes of labeled data. However, these systems are computationally expensive since they require pre-training followed by fine-tuning in a large parameter space. We explore the performance of such systems without fine-tuning by training a stateof- the-art speech recognizer on the fixed representations from the computationally demanding wav2vec 2.0 framework. We find performance to decrease without fine-tuning and, in the extreme low-resource setting, wav2vec 2.0 is inferior to its predecessor. In addition, we find that wav2vec 2.0 representations live in a low dimensional subspace and that decorrelating the features of the representations can stabilize training of the automatic speech recognizer. Finally, we propose a bidirectional extension to the original wav2vec framework that consistently improves performance.

AB - Recent advances in self-supervised learning through contrastive training have shown that it is possible to learn a competitive speech recognition system with as little as 10 minutes of labeled data. However, these systems are computationally expensive since they require pre-training followed by fine-tuning in a large parameter space. We explore the performance of such systems without fine-tuning by training a stateof- the-art speech recognizer on the fixed representations from the computationally demanding wav2vec 2.0 framework. We find performance to decrease without fine-tuning and, in the extreme low-resource setting, wav2vec 2.0 is inferior to its predecessor. In addition, we find that wav2vec 2.0 representations live in a low dimensional subspace and that decorrelating the features of the representations can stabilize training of the automatic speech recognizer. Finally, we propose a bidirectional extension to the original wav2vec framework that consistently improves performance.

KW - Automatic speech recognition

KW - Representation learning

KW - Self-supervised learning

KW - Semi-supervised learning

KW - Unsupervised learning

UR - http://www.scopus.com/inward/record.url?scp=85115116838&partnerID=8YFLogxK

U2 - 10.1109/ICASSP39728.2021.9414310

DO - 10.1109/ICASSP39728.2021.9414310

M3 - Article in proceedings

AN - SCOPUS:85115116838

VL - 2021-June

SP - 3885

EP - 3889

BT - ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

PB - IEEE

T2 - 2021 IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 2021

Y2 - 6 June 2021 through 11 June 2021

ER -