Parkinson's disease is one of the major neurodegenerative diseases that affect the central nervous system, often leading to motor and cognitive impairments in affected individuals. A precise diagnosis is currently unreliable, plus there are no specific tests such as electroencephalography or blood tests to diagnose the disease. Several studies have focused on the voice-based classification of Parkinson's disease. These studies attempt to enhance the accuracy of classification models. However, a major issue in predictive analysis is the imbalance in data distribution and the low performance of classification algorithms. This research aims to improve the accuracy of speech-based Parkinson's disease prediction by addressing class imbalance in the data and building an appropriate model. The proposed new model is to perform class balancing using SMOTE and build an ensemble voting model. The research process is systematically structured into multiple phases: data preprocessing, sampling, model development utilizing a voting ensemble approach, and performance evaluation. The model was tested using voice recording data from 31 people, where the data was taken from OpenML. The evaluation results were carried out using stratified cross-validation and showed good model performance. From the measurements taken, this study obtained an accuracy of 97.44%, with a precision of 97.95%, recall of 97.44%, and F1-Score of 97.56%. This study demonstrates that implementing the soft-voting ensemble-SMOTE method can enhance the model's predictive accuracy.