This paper introduces an iterative pseudo-labeling approach to improve automatic speech recognition for Mandarin-English code-switching, addressing the challenge of limited training data. The method leverages unlabeled corpora by generating pseudo-labels to create a semi-supervised dataset for two-stage bilingual model training.
The framework consists of three phases: pseudo-label generation from large unlabeled data, pre-training followed by fine-tuning on supervised code-switching data, and iterative refinements to enhance accuracy in complex scenarios.
This approach significantly advances code-switching ASR systems, achieving notable Mix Error Rate reductions of 6.35% on SEAME's devman subset and 8.29% on the devsge subset.