Parkinson’s disease affects millions of people worldwide, and changes in voice patterns are among the earliest detectable symptoms. Traditional machine learning methods for voice-based detection often struggle due to imbalanced datasets and suboptimal feature selection. To address these challenges, we propose a six-stage ensemble framework that combines advanced preprocessing, feature selection using an improved Zebra Optimization Algorithm (IZOA), and weighted ensemble learning. The preprocessing stage includes robust normalization and SMOTE-Tomek balancing, while IZOA optimizes both the selection of acoustic features and the ensemble classifier weights. Experiments were conducted on a dataset of 195 voice samples with 24 acoustic features, achieving an F1-macro score of 95.00%, a weighted F1-score of 95.00%, and an AUC of 99.38%, representing improvements of 7.85%, 3.06%, and 4.18%, respectively, over the best baseline methods. Analysis of feature importance highlights the contribution of shimmer, jitter, harmonics-to-noise ratio (HNR), and mel-frequency cepstral coefficients (MFCCs) to Parkinson’s voice characterization. The proposed approach provides a reliable and interpretable framework for early detection of Parkinson’s disease, demonstrating robustness and practical relevance for clinical applications.