Accurate and continuous heart rate (HR) monitoring is essential for health assessment and fitness tracking. Photoplethysmography (PPG) provides a low-cost, non-invasive sensing modality widely available in wearables, yet its reliability is limited by motion artifacts and inter-subject variability in free-living conditions. However, many prior approaches rely on multi-modal inputs or large models, and are often evaluated under subject-dependent pro-tocols, which do not adequately assess generalization performance in free-living conditions. Thus, we present a lightweight end-to-end model for heart-rate estimation from wrist PPG only, trained and evaluated under a strict leave-one-subject-out (LOSO) protocol, using the PPG-DaLiA dataset col-lected during daily activities. The architecture combines a compact 1D con-volutional front end with a bidirectional GRU back end, totaling approxi-mately 14.7k parameters and enabling embedded deployment. The proposed model achieves an MAE of 5.13 bpm, and a correlation of 0.86 across all LOSO test samples, showing strong agreement with ECG-derived reference and performance comparable to recent state-of-the-art methods. These find-ings indicate that accurate HR estimation is feasible with a single PPG chan-nel and compact architecture, supporting practical real-time deployment in wearable devices