Reproduce “Progressive Refinement: An Iterative Pseudo-Labeling Approach for Mandarin-English Code-Switching ASR” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qu Yang, Cakra Wardhana, Tim Ng
An iterative pseudo-labeling approach was developed to enhance Automatic Speech Recognition (ASR) for Mandarin-English code-switching speech by leveraging large amounts of unlabeled data. This method led to reductions in Mix Error Rate (MER) of 6.35% on SEAME devman and 8.29% on SEAME devsge, while maintaining strong performance on monolingual English speech.
Key insights
- Iterative pseudo-labeling effectively addresses the scarcity of labeled code-switching data by progressively refining ASR models through self-generated labels from unlabeled speech.
- A two-stage training strategy, involving pre-training on pseudo-labeled data followed by fine-tuning on supervised data, is crucial for optimal code-switching ASR performance, significantly outperforming single-stage or reversed-order training.
- Carefully tuned sampling weights for different data types (monolingual, code-switching, pseudo-labeled, supervised) during the fine-tuning stage are essential to balance generalization across languages and specific accuracy for code-switching segments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.