Speech-Based Parkinson's Disease Detection Using MFCC Features and Deep Representation Learning

Parkinson's disease (PD) is a progressive neurodegenerative disorder for which early diagnosis remains challenging, particularly in resource-constrained healthcare settings. Speech impairment is an early and prevalent symptom, motivating the development of non-invasive speech-based diagnostic approaches. This study proposes an image-based speech analysis framework for Parkinson's disease detection using spectrogram representations of speech signals. Speech recordings from the PC-GITA dataset, comprising 100 subjects (50 PD and 50 healthy controls), were segmented and transformed into time-frequency spectrogram images. Image preprocessing and data augmentation techniques were applied to improve robustness and generalization. Deep learning models were employed to automatically extract discriminative features from spectrogram images, which were subsequently classified using classical machine learning algorithms. Experimental results demonstrate that deep features extracted from spectrogram images (MFCC) using a Vision Transformer, combined with a Logistic Regression classifier, achieved an accuracy of 99%, indicating strong discriminative capability. The findings confirm that spectrogram-based speech representations can serve as reliable digital biomarkers for Parkinson's disease detection. The proposed framework offers a scalable, low-cost, and non-invasive screening solution suitable for real-world clinical deployment, particularly in resource-limited healthcare environments.