<p>Cardiovascular diseases (CVDs) remain the leading cause of mortality worldwide, highlighting the need for early, accurate, and accessible diagnostic methods. Phonocardiogram (PCG) signals contain valuable clinical information; however, their complex and non-stationary nature makes reliable interpretation challenging. Conventional machine learning approaches for heart sound classification often rely on cardiac cycle segmentation and handcrafted feature engineering, which can limit robustness and generalizability across diverse recording conditions. To address these limitations, this study proposes a deep audio embedding-based framework for the automatic classification of normal and abnormal heart sounds directly from unsegmented PCG recordings. Two pre-trained audio models, VGGish and YAMNet, originally developed for general audio recognition tasks, are employed to extract transferable and high-level audio embeddings. These embeddings are subsequently classified using conventional machine learning algorithms, including Support Vector Classifier (SVC), Random Forest (RF), XGBoost, Logistic Regression (LR), and K-Nearest Neighbors (KNN). Experimental evaluations were conducted on the PhysioNet CinC Challenge 2016 dataset using stratified five-fold cross-validation, with class imbalance addressed through the Synthetic Minority Oversampling Technique (SMOTE). The VGGish–SVC combination achieved the best performance, reaching an accuracy of 92.26% and a ROC-AUC of 96.67%, outperforming several existing state-of-the-art approaches. These results demonstrate that deep transfer learning with audio embeddings provides an efficient, robust, and clinically relevant solution for automated heart sound classification, supporting early cardiovascular disease screening and diagnosis.</p>