Voice analysis has emerged as a promising non-invasive biomarker for Parkinson's Disease (PD) diagnosis, owing to the prevalence of vocal impairments such as hypokinetic dysarthria in early disease stages. While existing approaches often rely on single-task phonation, they overlook the task-specific nature of vocal impairments in PD. To address this limitation, we propose a novel task-specific ensemble learning framework that models vocal recordings from multiple speech tasks. From each voice recording, we extract acoustic features from five clinically meaningful categories: jitter, shimmer, pitch, periodicity, and voice quality. An attention-based fusion module aggregates task-level predictions to produce interpretable and robust subject-level classifications. Experiments on a publicly available PD voice dataset demonstrate that our approach achieves 97.5% subject-level accuracy. Moreover, the attention weights offer valuable insights into the relative diagnostic importance of different vocal tasks. This work contributes a practical and interpretable voice-based PD screening system, suitable for deployment in telehealth and remote monitoring scenarios.