Convolutional Neural Network-based Classification of Parkinsonian Voice Disorders: A Spectrogram Approach

Parkinson’s disease produces a progressive degradation of motor (tremor, rigidity and slowness of movement) and non motor (cognitive, olfactory and speech) capabilities. Voice disorders are of great interest for patient monitoring and control, indeed as a possible predictor for early diagnosis of the disease. In this work, we implement three Deep Learning architectures to identify voice disorders associated with Parkinson’s disease from audio recordings. The audio records contain information on phonemes pronunciation and text reading. The proposed framework utilizes sliding windows to generate spectrograms for four ranges of frequencies. Then, the spectrograms are classified using a baseline convolutional neural network, ResNet50, and ResNet50V2. Results showed that the time of 2 seconds and spectral window range of $0-8000 \mathrm{~Hz}$ are the optimum settings to maximize the algorithm performance for both phonemes pronunciation and text reading. The experimental results showed that the proposed Resnet50v2 variant outperformed the other models, achieving a classification rate of $\mathbf{9 5 \%}$ for the phoneme classification, and $\mathbf{9 6. 6 7 \%}$ for the text classification. Results demonstrate that voice disorders in Parkinson’s disease can be accurately identified by the audio spectrograms using convolutional based architectures.