Towards Rehearsal-Free Multilingual ASR: A LoRA-based Case Study on Whisper
Tianyi Xu, Kaixun Huang, Pengcheng Guo, Yu Zhou, Longtao Huang, Hui Xue, Lei Xie
TL;DR
This work tackles the problem of adapting large multilingual ASR models like Whisper to low-resource languages without access to the original training data, addressing catastrophic forgetting. It introduces orthogonal gradient-based continual learning with LoRA updates, via O-LoRA and O-AdaLoRA, leveraging the original model's LoRA subspace and a learnable rank mechanism to achieve rehearsal-free, parameter-efficient fine-tuning. The approach is formalized and evaluated on Chinese, Uyghur, and Tibetan data, showing reduced forgetting and improved convergence compared to baselines, while using a compact parameter budget. The findings have practical implications for scalable, privacy-preserving deployment of speech foundation models to underrepresented languages, enabling robust multilingual ASR without retraining on historical data.
Abstract
Pre-trained multilingual speech foundation models, like Whisper, have shown impressive performance across different languages. However, adapting these models to new or specific languages is computationally extensive and faces catastrophic forgetting problems. Addressing these issues, our study investigates strategies to enhance the model on new languages in the absence of original training data, while also preserving the established performance on the original languages. Specifically, we first compare various LoRA-based methods to find out their vulnerability to forgetting. To mitigate this issue, we propose to leverage the LoRA parameters from the original model for approximate orthogonal gradient descent on the new samples. Additionally, we also introduce a learnable rank coefficient to allocate trainable parameters for more efficient training. Our experiments with a Chinese Whisper model (for Uyghur and Tibetan) yield better results with a more compact parameter set.