Abstract
This study investigates transformer fine-tuning strategies for low-resource dysarthric speech recognition across four languages (English, Polish, Italian, and Dutch). We compare full-parameter fine-tuning, Low-Rank Adaptation (LoRA), and a QK-MLP ablation that updates the same layers as LoRA but uses full-rank parameter updates, alongside regularisation techniques such as dropout and SpecAugment. LoRA outperforms full-parameter fine-tuning on five of six test sets, achieving new state-of-the-art word error rates (WER) on COPAS (23.86 vs. 29.0), EasyCall (17.15 vs. 39.2), and PLDD (69.59 vs. 71.91), and providing the first large-scale ASR evaluation on GidoLab. A direct comparison with the QK-MLP variant confirms that LoRA's advantage arises from its low-rank constraint acting as implicit regularisation, rather than simply from the selection of updated layers. Severity-stratified evaluation further shows that LoRA reduces cross-severity WER spread, indicating improved robustness to dysarthria severity variation and suggesting that the low-rank constraint prevents severity-specific overfitting.
Paper Type
Full Paper
DOI
10.62036/ISD.2026.94
Robust Low-Resource Dysarthric Speech Recognition via Low-Rank Adaptation
This study investigates transformer fine-tuning strategies for low-resource dysarthric speech recognition across four languages (English, Polish, Italian, and Dutch). We compare full-parameter fine-tuning, Low-Rank Adaptation (LoRA), and a QK-MLP ablation that updates the same layers as LoRA but uses full-rank parameter updates, alongside regularisation techniques such as dropout and SpecAugment. LoRA outperforms full-parameter fine-tuning on five of six test sets, achieving new state-of-the-art word error rates (WER) on COPAS (23.86 vs. 29.0), EasyCall (17.15 vs. 39.2), and PLDD (69.59 vs. 71.91), and providing the first large-scale ASR evaluation on GidoLab. A direct comparison with the QK-MLP variant confirms that LoRA's advantage arises from its low-rank constraint acting as implicit regularisation, rather than simply from the selection of updated layers. Severity-stratified evaluation further shows that LoRA reduces cross-severity WER spread, indicating improved robustness to dysarthria severity variation and suggesting that the low-rank constraint prevents severity-specific overfitting.
Recommended Citation
Piernicki, T. & Kostek, B.(2026). Robust Low-Resource Dysarthric Speech Recognition via Low-Rank Adaptation. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.94