Speaker
Description
In this study, we evaluate a recently proposed hypernetwork‑based, parameter‑efficient fine‑tuning approach for dysarthric speech recognition in Hungarian and English, using the Whisper ASR (Automatic Speech Recognition) model. We compare its performance to Low‑Rank Adaptation (LoRA) demonstrating the promise of hypernetwork‑driven personalization for atypical ASR. To better understand speaker‑level variability, we analyze the encoder’s latent‑space representations and examine the presence of intra‑cohort structure. Our results show that, even under severe data limitations, the hypernetwork‑based method consistently improves recognition accuracy for previously unseen speakers while requiring substantially fewer trainable parameters than LoRA, yet achieving competitive performance. These findings highlight the potential of hypernetwork‑based adaptation as an efficient and effective strategy for dysarthric ASR in low‑resource scenarios.