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

TL;DR

This study examines how self-supervised speech pre-training transfers to Flemish Dutch by comparing multiple pre-training methods (APC, Mockingjay, CPC, wav2vec, wav2vec 2.0, and XLSR-53) across varying data regimes. It demonstrates that both a large amount of pre-training data and matching the pre-training domain (language and speech type) are key for positive transfer, with finetuning on Flemish transcripts delivering substantial gains. Wav2vec 2.0, especially the large and multilingual XLSR-53 variants, yields the strongest improvements, achieving up to a ~30% relative WER reduction when fine-tuned on Flemish and evaluated with an HMM-DNN acoustic model. The results highlight that cross-lingual transfer is feasible for Flemish Dutch, and that domain-aligned pre-training plus targeted finetuning can meaningfully reduce WER in real-world ASR settings, even under noisy conditions when augmented finetuning data are used.

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.