🤖 AI Summary
Whisper’s small variants suffer substantial performance degradation on low-resource language automatic speech recognition (ASR) and struggle to jointly optimize multilingual capability and robustness. To address this, we propose a lightweight multilingual knowledge distillation framework featuring a novel dual-path mechanism: (1) language-specific expert fine-tuning (LoE) via modular ASR adaptation on Whisper-small, incorporating language-adaptive LoE adapters; and (2) cross-task, cross-lingual knowledge distillation from Whisper-large-v2. We further integrate multi-domain data augmentation. Experiments demonstrate consistent and significant WER reductions over standard fine-tuning and LoRA on both in-domain and out-of-domain test sets for target low-resource languages. The approach incurs negligible parameter overhead (<0.5% additional parameters), effectively bridging the resource gap while fully preserving Whisper’s inherent multilingual and multitask generalization capabilities.
📝 Abstract
Whisper is a multitask and multilingual speech model covering 99 languages. It yields commendable automatic speech recognition (ASR) results in a subset of its covered languages, but the model still underperforms on a non-negligible number of under-represented languages, a problem exacerbated in smaller model versions. In this work, we propose DistilWhisper, an approach able to bridge the performance gap in ASR for these languages while retaining the advantages of multitask and multilingual capabilities. Our approach involves two key strategies: lightweight modular ASR fine-tuning of whisper-small using language-specific experts, and knowledge distillation from whisper-large-v2. This dual approach allows us to effectively boost ASR performance while keeping the robustness inherited from the multitask and multilingual pre-training. Results demonstrate that our approach is more effective than standard fine-tuning or LoRA adapters, boosting performance in the targeted languages for both in- and out-of-domain test sets, while introducing only a negligible parameter overhead at inference.
