Heart rate variability is a well-established noninvasive marker of autonomic regulation, yet robust automatic recognition of rest, fatigue, and stress in field conditions remains challenging. We present a hybrid pipeline that combines generative and discriminative evidence with temporal smoothing. First, HRV windows (meanRR, SDNN, RMSSD, pNN50, SampEn, DFA α1) are standardized and modeled by a Gaussian Mixture Model to obtain soft memberships. In parallel, an SVM (RBF) trained with weak labels (per condition) produces calibrated posteriors. The two probability streams are fused by a weighted sum, and a Hidden Markov Model with physiologically plausible transitions (favoring Res vs. Fatigu vs. Stress over direct Res vs. Stress) yields the final Viterbi sequence. Evaluation follows a leave-one-file-out protocol, where the proposed hybrid method achieves Accuracy 0.94, Macro-F1 0.92, ARI 0.85, and NMI 0.88, clearly outperforming K-Means (Accuracy 0.86, Macro-F1 0.87) and standalone GMM (Accuracy 0.83, Macro-F1 0.84). Per-class results show balanced performance with F1 0.93 for Rest, 0.91 for Fatigue, and 0.93 for Stress, substantially reducing Rest vs. Stress misclassifications and improving Fatigue recognition. PCA visualizations support geometric separability, while DBSCAN on PC1–PC2 acts as an outlier detector for artifact-prone windows. Across experiments, the hybrid consistently improves Macro-F1 by 5–15 percentage points over the best single baseline (K-Means, standalone GMM/HMM), providing more stable temporal trajectories without sacrificing interpretability. The method is lightweight and amenable to near real-time deployment on wearable ECG/PPG devices, enabling practical monitoring and prevention of accumulated fatigue and acute stress in athletes.