Voice-Specific Augmentations for Parkinson’s Disease Detection Using Deep Convolutional Neural Network

Deep convolutional neural networks (CNN) have provided remarkable results in speech processing tasks. However, they require a large amount of training data to prevent overfitting. Data augmentation is commonly used to increase the quantity of training data by altering the original samples. Furthermore, it improves the generalization and robustness of CNN models. This work provides a deep learning approach for detecting Parkinson’s disease (PD) from voice recordings converted to spectrograms. We investigate the use of six voice-specific data augmentation techniques in PD detection to improve generalization to unseen data. We utilized the Xception model to evaluate the performance of the proposed approach. We separately analyzed the articulatory and phonatory aspects of the voice of PD patients by considering recordings of vowels and words from the PC-GITA dataset. The best results were achieved on vowel data. Our model yielded 92.33% accuracy in the case of the /a/ vowel subset. In the case of words, the model achieved 82.12% accuracy on the whole words set. The obtained results show that voice-specific augmentation techniques improve the detection of speech impairments in PD patients.