A controlled comparative evaluation of data augmentation strategies for low-resource dysarthric speaker recognition

Life Cycle Reliability and Safety Engineering · Springer Nature · 2026-09-30 · DOI: 10.1007/s10772-026-10289-2

Аннотация

Developing effective Dysarthric Speaker Recognition (DSR) systems remains challenging due to the limited availability of pathological speech data and the high intra-speaker variability caused by articulatory impairments. This study presents a systematic comparative evaluation of three data augmentation strategies: conventional waveform-based augmentation, feature-level masking using SpecAugment, and synthetic data generation with conditional Generative Adversarial Networks (cGANs), all integrated into a unified two-dimensional Convolutional Neural Network (2D-CNN) framework for closed-set DSR. Rather than introducing a new augmentation technique, the objective is to assess the relative effectiveness of existing approaches under identical experimental conditions. Augmentation was applied exclusively to the training data using strict utterance-disjoint train-test separation, while testing was performed on original, non-augmented utterances. To ensure robust evaluation and reduce partition bias, five independent random utterance-level splits were employed. All augmentation strategies consistently improved recognition performance over the non-augmented baseline. SpecAugment achieved the highest average accuracy and the most stable performance across all splits, whereas cGAN-based augmentation generated more diverse training samples but showed greater optimization variability. Although the performance difference between SpecAugment and cGAN was not statistically significant across repeated evaluations, SpecAugment demonstrated superior consistency. Furthermore, severity-based analysis revealed that data augmentation is particularly beneficial for speakers with severe dysarthria. These findings provide statistically validated evidence that feature-level augmentation offers a reliable solution for low-resource dysarthric speaker recognition, while generative augmentation remains a promising approach for modeling pathological speech variability in clinical speech technology applications.

Читать у издателя →

Похожие статьи