Parkinson’s disease (PD) is a progressive disorder that affects movement and speech, making early diagnosis important to improve patient care and quality of life. In this work, we developed and compared several machine learning models for detecting PD based on voice recordings. The speech samples used in the study came from two different public datasets and included both people with Parkinson’s and healthy individuals. From these recordings, we extracted 159 features capturing temporal, spectral, and cepstral information, such as MFCCs and Bark-scale energies. To reduce the number of features and enhance model interpretability, we applied two different selection methods: Random Forest feature importance and linear SVM with L1 regularization. This resulted in two smaller sets containing 52 and 40 features, respectively. Using these reduced sets, we trained and tested a range of classifiers including Random Forest, SVM (RBF), k-nearest neighbors, logistic regression, and a simple neural network. Our best results were obtained with an SVM using the SVM L1-selected features, reaching an accuracy and F1-score of 86% in the test set. These findings show that careful feature selection can not only simplify models but also improve the accuracy of automatic PD detection from voice. Comparing multiple selection strategies and classifier types proved to be a key factor for optimizing the system’s performance.