<p>Parkinson’s disease is a progressive neurodegenerative disorder that affects about 1% of the global population over the age of 55. It presents with a range of motor and non-motor symptoms, with voice disorders often being one of the earliest non-motor indicators. Early diagnosis of Parkinson’s can be facilitated by analyzing voice recordings from patients. This study aims to develop differential diagnosis models to distinguish between Parkinson’s patients and healthy individuals using audio features. The methodology involves training and classifying data with one-dimensional convolution, two-dimensional convolution, and convolutional autoencoder networks. The study dataset comprises voice recordings from 188 Parkinson’s and 64 healthy individuals, resulting in 756 voice samples and the extraction of 754 audio features. Key feature sets, such as wavelet transform (WT), tunable Q-factor wavelet transform (TQWT), and mel frequency cepstral coefficient (MFCC), along with their combinations, were used as input data. The results revealed that the highest classification accuracies were 0.90% and 0.83% for one-dimensional convolution and convolutional-auto-encoder networks using the TQWT feature, respectively. The two-dimensional network achieved a top accuracy of 0.86% with a combination of WT and MFCC features. Among the machine learning classifiers, average accuracies were 0.75% for support vector machine, 0.74% for nearest neighbor, 0.78% for logical regression, 0.85% for random forest, 0.52% for stochastic gradient descent, and 0.34% for multilayer perceptron methods. In conclusion, the one-dimensional network demonstrated superior performance over the other networks, and the random forest classifier outperformed other machine learning methods. Additionally, TQWT was identified as the most effective diagnostic feature.</p>