Various approaches of spectrogram for CNN-based classification of Parkinson's and healthy speech

Parkinson's disease (PD) is a prevalent neurodegenerative disorder on a global scale. The present method for identifying Parkinson's disease is costly and time-consuming. In recent times, convolutional neural networks (CNNs), a deep learning technique have gained ground in automatic classification because of their potential to extract complex nonlinear patterns from input data. This work compares the accuracy of the CNN model for the spectrogram in DB and non-DB scales and divides the obtained spectrograms into low-frequency and high-frequency spectrograms in both scales. This is achievable because the convolution layer in CNN automatically gains knowledge of the relevant features from the spectrograms. In this work, our model achieves a competitive classification accuracy of 100% using a low-frequency spectrogram in a non-DB scale which is half the frequency component of the actual spectrogram which demonstrates the effectiveness of the low-frequency component of the spectrogram in distinguishing Parkinson's speech and healthy controls. This study spotlights the capacity of deep learning in creating non-intrusive, automated algorithms for the early identification and investigation of Parkinson's disease, therefore facilitating the development of more easily accessible and precise diagnostic instruments.