Parameter-efficient Taiwanese Hokkien speech intent classification using Whisper and low-rank adaptation (LoRA)

Taiwanese Hokkien is a low-resource language that remains underexplored in speech‑understanding tasks. Although large-scale pre-trained models such as Whisper have shown strong multilingual capabilities, adapting them to low-resource settings is computationally expensive when using full fine-tuning. To address this limitation, this study proposes a parameter-efficient speech intent classification framework based on Whisper-small with Low-Rank Adaptation (LoRA). The proposed method is evaluated on a Taiwanese Hokkien speech intent dataset consisting of 1,600 training, 519 validation, and 960 test utterances across eight intent classes. Extensive experiments, including convergence analysis, five-run statistical evaluation, LoRA rank ablation studies, comprehensive baseline comparisons, intent-level performance analysis, and confusion matrix analysis, validate the proposed framework. It achieves an average accuracy of 89.17 ± 1.64% and a Macro F1-score of 89.19 ± 1.64%. Among all configurations, LoRA rank 32 achieves the highest average performance among the tested configurations between performance and efficiency. The proposed approach also consistently outperforms AdaLoRA, Frozen Encoder, Whisper-base with LoRA, and Full Fine-Tuning while updating only 1.18 million parameters (1.32%) of the Whisper-small model. These results demonstrate that parameter-efficient adaptation through LoRA provides an effective and computationally efficient solution for low-resource Taiwanese Hokkien speech intent classification.