Recent advances position speech analysis as a promising non-invasive biomarker for Parkinson’s disease (PD), with deep learning (DL) models achieving notable classification performance. However, the inherent complexity of such models underscores the need for robust explainability to ensure clinical interpretability, trust, and real-world applicability. To address this limitation, we extract saliency maps from Transformer-based models to visualize key time-frequency regions, thereby identifying clinically meaningful features for PD diagnosis. Specifically, we evaluate CTNets, which integrate convolutional layers for localized spectro-temporal feature extraction with Transformer-based attention mechanisms to capture long-range dependencies. To enhance model explainability, we assess three activation mapping techniques: Layer-wise Relevance Propagation, Gradient-weighted Class Activation Mapping, and Score-weighted Class Activation Mapping. Validation across two real-world datasets reveals a high degree of consistency in attention patterns, underscoring CTNet’s capacity to combine advanced ML methodologies with clinically relevant neurological assessment. By ensuring interpretability and demonstrating robust generalization across multilingual clinical datasets, the CTNet framework establishes a reliable foundation for computer-aided PD detection.