Adaptive Accented Speech Recognition via Feature Disentanglement and Meta-learning

Automatic speech recognition (ASR) systems have achieved substantial progress with deep learning techniques; however, accurately recognizing accented speech remains challenging due to inherent accent variability, limited accent-specific data, and high acquisition costs. To address these limitations, we propose an adaptive accented speech recognition framework leveraging feature disentanglement through multi-task learning, coupled with meta-learning techniques. Specifically, we design a Transformer-based adaptive multi-task architecture to simultaneously perform accented speech recognition and accent classification, effectively disentangling accent-specific information from accent-invariant semantic features. Furthermore, we introduce a Model-Agnostic Meta-Learning (MAML) strategy to rapidly adapt the ASR model to unseen accents, significantly improving robustness and generalization capabilities. Extensive experiments conducted under both mixed-regional and cross-regional scenarios demonstrate that our method consistently outperforms conventional joint training and standard MAML methods, particularly under conditions involving scarce training data and previously unseen accent classes. Our approach thus provides a practical and efficient solution for accented speech recognition tasks in real-world scenarios.