EMGL: Enhancing Mandarin-English Code-Switching Speech Recognition using Generative Labeling Technique

Mandarin-English Code-Switching Automatic Speech Recognition (MECS-ASR) is significantly impeded by limited data availability. Previous studies have demonstrated only marginal improvements in performance, even with the implementation of data augmentation, pre-trained models, prompt engineering, and multi-task learning strategies. This paper presents a groundbreaking approach, Enhancing MECS-ASR using Generative Labeling Technique (EMGL), which integrates text generation, forced alignment, speech synthesis, and verification via ASR technologies. Our method has synthesized a substantial corpus of approximately 1200 hours of Mandarin-English mixed speech, significantly expanding the available data. Experiments on a randomly initialized model demonstrate that the EMGL-generated dataset substantially improves MECS-ASR accuracy. This improvement is achieved without the need for additional language classifiers or refined prompt engineering techniques. Our results highlight the transformative potential of EMGL in addressing the persistent challenge of data scarcity in MECS-ASR, offering a robust solution to a field in need of advancement.