Scaling Pre-training Data Volume Effects on Speech Recognition Performance in Flemish Dutch

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

Research goal: How does scaling the pre-training data volume from 100 to 1000 hours affect the downstream speech recognition performance of contrastive, masked prediction, and latent diffusion pre-trained models on Flemish Dutch, measured by phoneme error rate and computational efficiency during fine-tuning?

Autonomous synthesis report generated by Assignee Research. Tribunal consensus score: 7.5/10.