Ensuring poultry health is essential for global food security and rural livelihoods, particularly in low- and middle-income regions where poultry farming serves as a vital financial buffer. Traditional health monitoring techniques, such as visual inspections and laboratory diagnostics, are labor-intensive and poorly scalable for large flocks. To address these challenges, this study proposes a hybrid deep learning framework that combines Mel-spectrogram based audio feature extraction with a Swin Transformer and a Random Forest classifier optimized through Harmony Search. The Poultry Vocalization Signal Dataset, comprising healthy, unhealthy, and noise clips recorded at 96 kHz, was preprocessed using advanced noise reduction and dual feature extraction pipelines, producing both handcrafted statistical features and Mel-spectrogram images. The Swin Transformer efficiently captures long-range spectral dependencies, while the Random Forest provides lightweight, interpretable decision-making. Experimental results demonstrate that ST-HRF achieves a test accuracy of 98.12%, with a Matthews Correlation Coefficient (MCC) of 0.9784, Critical Success Index (CSI) of 0.9612, F1-score of 0.9727, and Cohen’s Kappa of 0.9596, significantly outperforming conventional CNN architectures and alternative optimizers. Furthermore, detailed evaluations using ROC curves, precision-recall analysis, cumulative gain, lift charts, calibration, and entropy plots confirm the model’s robustness, reliability, and strong generalization. The proposed framework offers a scalable, efficient, and explainable solution for continuous on-farm poultry health surveillance, paving the way for real-world deployment in resource-constrained agricultural environments.