On Scaling Contrastive Representations for Low-Resource Speech Recognition

TL;DR

This paper investigates the use of fixed contrastive speech representations for low-resource speech recognition, revealing limitations without fine-tuning and proposing a bidirectional extension to improve performance.

Contribution

It demonstrates the challenges of using fixed wav2vec 2.0 representations without fine-tuning and introduces a bidirectional extension that enhances recognition accuracy.

Findings

Performance drops without fine-tuning, especially in low-resource settings.

Wav2vec 2.0 representations are low-dimensional and decorrelation stabilizes training.

Bidirectional extension improves speech recognition performance.

Abstract

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 state-of-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.