Research report: Comparative analysis of GAN-based augmentation pipelines for Parkinson’s hypokinetic dysarthria severity detection using deep learning

Class imbalance is one of the quiet saboteurs of medical machine learning: when a dataset contains far more examples of one category than another, models tend to favor the majority class and overlook the rarer, often more clinically important cases. In speech-based assessment of Parkinson’s disease, this problem is acute, because recordings of severely affected speakers are scarce. A new study published in Frontiers in Artificial Intelligence tackles the problem with generative adversarial networks, or GANs, and asks a deceptively simple methodological question: should synthetic training data be created as artificial speech waveforms, or directly as artificial images of the speech signal?

The research, conducted by Aurobindo S and Prakash Ramachandran of Vellore Institute of Technology in India and Rajesh Kumar M of VIT Mauritius, compared two complete augmentation pipelines for grading hypokinetic dysarthria, the soft, monotone, imprecise speech disorder common in Parkinson’s disease. The first pipeline, called GAN-generated speech (GGS), trained a deep convolutional GAN on log-magnitude spectrograms of pathological sustained vowels, generated synthetic spectrograms, reconstructed waveforms from them using iterative phase estimation, and then converted the synthetic audio into low-frequency cepstrogram images. The second pipeline, called GAN-generated image (GGI), skipped the audio stage entirely: a Wasserstein GAN with gradient penalty (WGAN-GP) was trained directly on low-frequency cepstrogram images of the minority class to produce synthetic images.

The choice of representation was not arbitrary. The low-frequency cepstrogram, which retains the first 128 cepstral bins of each short-time frame, had previously outperformed other time-frequency representations in the same research group’s work on dysarthria detection and severity classification, and the low-quefrency cepstral coefficients mainly encode slow-varying vocal-tract and spectral-envelope information relevant to pathological voice. All recordings came from the PC-GITA corpus, a Spanish dataset of 50 Parkinson’s patients and 50 age- and gender-matched healthy controls. Only the patients’ sustained vowels were used, since the task was to distinguish low severity, defined as a UPDRS speech score of 0 to 1, from high severity, defined as a score of 2 to 3, among patients rather than to separate patients from controls. The original dataset contained 300 high-severity and 450 low-severity samples, with each speaker contributing three repetitions of each of the five Spanish vowels.

Both pipelines generated 150 synthetic high-severity samples, raising the minority class to parity at 450 samples, and the synthetic data were used exclusively for training. The downstream classifier was the same in every condition: a four-block convolutional neural network with batch normalization, ReLU activations, max-pooling, dropout, and a softmax output, fed with 128 by 128 pixel cepstrogram images. The authors are explicit that the two GAN pipelines were not a strictly controlled domain-only comparison. The speech-domain GAN used a conventional adversarial objective at a 128 by 640 resolution, while the image-domain WGAN-GP worked at 128 by 128 with a Wasserstein objective and gradient penalty, so performance differences may reflect the GAN formulation and training configuration as well as the augmentation stage.

The evaluation was run under two protocols. In the speaker-independent setting, all 15 utterances of each speaker stayed within a single partition, so no speaker appeared in both training and testing; this is the harder and more clinically meaningful test of generalization. In the speaker-dependent setting, utterances from the same speaker could appear in training, validation, and testing, which the authors candidly describe as an optimistic upper bound, since the model can exploit speaker identity, and part of the observed accuracy may reflect speaker recognition rather than severity classification.

In the single-split speaker-independent experiment, the original imbalanced baseline achieved a test accuracy of only 48.33 percent, slightly below the 50 percent chance level of the balanced test set, with a large gap between training accuracy of 81.25 percent and validation accuracy of 58.67 percent. GGS augmentation lifted test accuracy to 59.17 percent, precision to 60.78 percent, and specificity to 66.67 percent, while recall stayed at 51.67 percent. GGI performed better still, reaching 61.67 percent test accuracy, 63.46 percent precision, 55.00 percent recall, 68.33 percent specificity, and an F1-score of 58.93 percent, exceeding GGS on every reported sample-level metric in this split.

Because a single split can be misleading, the team repeated five-fold speaker-stratified cross-validation with five independent random seeds, yielding 25 fold-seed evaluations per condition. Here the picture shifted: GGS achieved the highest mean accuracy at 56.47 plus or minus 3.40 percent, compared with 55.47 plus or minus 3.35 percent for GGI and 53.13 plus or minus 3.26 percent for the original baseline. GGS also led in mean precision and specificity, whereas GGI achieved higher mean recall and a marginally higher mean F1-score. The authors stress that the improvement over baseline was small relative to the variability across folds and seeds, and that the repeated analysis was a robustness check rather than a fully nested evaluation in which GAN training is repeated inside every fold. McNemar’s exact tests on pooled paired predictions showed significant changes in classification decisions for both augmented conditions, but per-fold paired t-tests, which respect the dependence of repeated predictions, found significant accuracy gains for GGS in three of five folds and for GGI in only one fold.

Under the speaker-dependent protocol, GGI dominated. In the single split it reached 73.13 percent test accuracy with 80.39 percent precision, 61.19 percent recall, 85.07 percent specificity, and an F1-score of 69.49 percent, cutting false positives from 22 under GGS to just 10. GGS, by contrast, scored 63.43 percent, below the original baseline’s 66.42 percent in that split. Across the repeated speaker-dependent evaluations, GGI achieved a mean accuracy of 76.48 plus or minus 4.09 percent, versus 73.04 plus or minus 3.63 percent for GGS and 73.04 plus or minus 3.25 percent for the baseline. McNemar testing showed a significant paired difference for GGI against its baseline and none for GGS, and per-fold t-tests confirmed significant GGI gains in three of five folds. Supplementary comparisons with class-weighted cross-entropy, SMOTE, and SpecAugment-style masking showed that those conventional strategies produced distinct trade-offs, with SpecAugment boosting recall at the cost of specificity, while the GAN pipelines achieved higher accuracy in the fixed evaluations.

The authors also probed how similar the synthetic data were to the real data. A t-SNE projection showed substantial overlap between GGI images and real cepstrograms, while GGS samples spread more widely. Kernel Inception Distance was 0.197448 for GGS and 0.263333 for GGI, though the authors caution that these values were interpreted separately within each pipeline and not compared directly, because the representations differ and Inception features are not calibrated for non-natural images. Nearest-neighbor analysis found that GGI samples sat closer to individual real samples but drew on a smaller subset of unique real neighbors, whereas GGS showed greater diversity in its nearest real samples. The team also checked whether utterance duration could explain the GGS results: a duration-only logistic regression was not statistically significant, and post-processed synthetic samples were on average shorter than real high-severity recordings, although duration-related characteristics could still be implicitly encoded in the cepstrograms.

The study’s absolute accuracies, generally in the 50 to 76 percent range, are far lower than the above-95 percent figures reported for multi-class dysarthria severity on continuous-speech corpora such as UA-Speech and TORGO, and the authors attribute this to the binary severity boundary drawn within a continuous clinical scale, the sustained-vowel task that limits prosodic and rhythmic cues, and the small cohort. Their conclusions are appropriately measured: GAN-based augmentation showed potential for addressing class imbalance under the conditions studied, but no strategy was consistently superior, the relative performance of GGS and GGI depended on the evaluation protocol and the metric, and the speaker-dependent advantage for GGI should be read as an upper bound rather than evidence of generalization. Objective perceptual validation of the reconstructed speech, for example with PESQ or STOI, was not performed, and controlled experiments using the same GAN formulation in both domains remain necessary to isolate the true effect of augmenting speech versus images.

Subject of Research: Comparing GAN-generated speech and GAN-generated image augmentation for class imbalance in Parkinson's hypokinetic dysarthria severity detection

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

Article References: S, A., Ramachandran, P., & M, R. K. (2026). Comparative analysis of GAN-based augmentation pipelines for Parkinson's hypokinetic dysarthria severity detection using deep learning. Frontiers in Artificial Intelligence, 9, Article 1951063. https://doi.org/10.3389/frai.2026.1951063

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1951063

Keywords: Parkinson's disease, hypokinetic dysarthria, generative adversarial networks, data augmentation, class imbalance, convolutional neural network, cepstrogram, PC-GITA dataset, speaker-independent evaluation, speech processing, severity classification, WGAN-GP