Comparative analysis of GAN-based augmentation pipelines for Parkinson's hypokinetic dysarthria severity detection using deep learning

Class imbalance is a common problem in speech-based severity detection, which can result in biased models and poor recognition of minority severity classes. Generative Adversarial Networks (GANs) have been widely used for data augmentation, but very few studies have comparatively evaluated GAN-based augmentation at different stages of the speech-processing pipeline for assessing Parkinson's hypokinetic dysarthria severity. This paper compares GAN-generated speech (GGS) and GAN-generated image (GGI) augmentation methods for addressing class imbalance in low-frequency cepstrogram-based severity detection using the PC-GITA dataset. GAN-based augmentation was used to generate balanced training datasets using two different augmentation pipelines. The original speech and synthetic speech generated using GGS were converted into low-frequency cepstrogram representations, whereas GGI directly generated low-frequency cepstrogram images. These representations were used as inputs to a CNN trained and evaluated under Speaker-independent (SI) and Speaker-dependent (SD) conditions. In the speaker-independent evaluation, repeated five-fold speaker-stratified cross-validation with five independent random seeds yielded a mean accuracy of 56.47% ± 3.40% for GGS compared to 55.47% ± 3.35% for GGI and 53.13% ± 3.26% for the original-dataset baseline. GGS showed higher mean accuracy, precision, and specificity, whereas GGI showed higher mean recall and F1-score. In the speaker-dependent evaluation, GGI achieved the highest mean accuracy of 76.48% ± 4.09% compared to 73.04% ± 3.63% for GGS and 73.04% ± 3.25% for the original-dataset baseline. However, the speaker-dependent results represent an optimistic upper-bound evaluation because speaker identity was shared between the training and evaluation partitions. In general, the two GAN-based augmentation pipelines showed potential to address class imbalance under the conditions studied, while their relative performance depended on the evaluation protocol and the classification measure.