Fetal ECG (fECG) signal helps in identifying fetal heart abnormalities along with fetal health during pregnancy, which reduces neonatal mortality and postnatal complications. During invasive procedure, its acquisition is risky as the electrode is placed over the fetus head, which may lead to contamination and infection to both mother and fetus. In few cases, fluid may come out during the placement of electrode and makes the health condition of baby to be critical. But, researchers used direct fECG as reference for separating it from abdominal ECG (aECG). So, electrodes are placed over abdomen of the expecting mother for capturing fECG signal. Non-invasive procedure of capturing fECG recordings from the mother’s abdomen is convenient, but its extraction is challenging due to its contamination with maternal ECG (mECG), noise, and its lower magnitude compared to mECG. Due to the fetal movement, sometimes its acquisition is not accurate. Researchers used direct fECG (captured from the fetal head) as reference for extracting fECG from aECG. But, direct fECG is limited in the database for making the deep learning (DL) models to perform better. In this paper, generative adversarial network (GAN) is used for obtaining synthetic fECG, which may help extracting mECG and fECG from real aECG. These deep neural networks are capable of generating the signals identical in distribution, shape, and information content by using generator and discriminator. In this manuscript, generator and discriminator block consist of 5 numbers of transposed convolution and convolution layers, respectively. The deviation between generated fECG and real fECG are measured by mean, variance, and RMS. In this work, direct fECG and aECG available in ADFECG database have been used for generating fECG and aECG. The generated signals have been used by SVM classifier for classification and reported an overall accuracy of 93.75 \(\%\) .