Comparison of Self-Supervised Speech Pre-Training Methods on Flemish Dutch

TL;DR

This study compares various self-supervised speech pre-training methods on Flemish Dutch, highlighting the importance of data quantity and domain matching for effective transfer, with wav2vec 2.0 showing superior results.

Contribution

It provides a comprehensive comparison of pre-training methods on Flemish Dutch, emphasizing the impact of data size and domain similarity on downstream speech recognition performance.

Findings

Pre-trained models improve linear phone separability.

Not all methods enhance Automatic Speech Recognition.

Finetuning XLSR-53 yields 30% WER reduction.

Abstract

Recent research in speech processing exhibits a growing interest in unsupervised and self-supervised representation learning from unlabelled data to alleviate the need for large amounts of annotated data. We investigate several popular pre-training methods and apply them to Flemish Dutch. We compare off-the-shelf English pre-trained models to models trained on an increasing amount of Flemish data. We find that the most important factors for positive transfer to downstream speech recognition tasks include a substantial amount of data and a matching pre-training domain. Ideally, we also finetune on an annotated subset in the target language. All pre-trained models improve linear phone separability in Flemish, but not all methods improve Automatic Speech Recognition. We experience superior performance with wav2vec 2.0 and we obtain a 30% WER improvement by finetuning the multilingually pre-trained XLSR-53 model on Flemish Dutch, after integration into an HMM-DNN acoustic model.