Introduction: Assessment of fetal health during pregnancy is essential to ensure the delivery of a healthy offspring. However, conventional clinical tools including echocardiography are still heavily dependent on expert interpretation which makes it a difficult task in lengthy and frequent sessions. Methods: Here, we propose for the first time the use of a deep learning-based approach using U-Net neural networks for the extraction of fetal phonocardiograms (PCG) and interpretation of fetal heart rates as a reflection of fetal well-being. A total of 20 healthy pregnant women were included in the study and asked to record 4-channel PCG for 10 minutes. Data preparation included only the identification of less-noisy channels and the removal of low-frequency breathing sound waves. The U-Net model was then trained using 1-second PCG segments and validated through a leave-one-subject-out (LOSO) cross-validation scheme to predict patient-wise fetal heart activity. The performance was evaluated relative to ground-truth fetal PCG identified through fetal electrocardiography (ECG) masks. Results: The model successfully extracted fetal PCG with a median root mean square error (RMSE) of 0.702 [IQR: 0.695-0.706] compared to the ground-truth. Moreover, the estimation of fetal heart rates using the U-Net-based fetal PCG had a correlation with the ground-truth of 0.642 (p-value = 0.002) and a median error of 18.508 [IQR: 11.996-23.215]. The Bland-altman analysis revealed a slight decrease in mean heart rates using the U-Net approach of 5.18±24.97. Conclusion: This study suggests deep learning as an automated approach to extract fetal heart sounds and accurately diagnose fetal well-being. It could act as a clinical AI-assisted tool to clinicians to reduce the dependency on medical experts when continuous and frequent assessment sessions are required.