Detecting earthquake events in seismic time series can be a daunting task. Traditional human-based visual detection has been the established benchmark, but it demands substantial manual effort and doesn't efficiently adapt to large datasets. Over the past few years, machine learning detection techniques have been adopted to enhance accuracy and efficiency. However, the effectiveness of these approaches hinges on the availability of a substantial volume of high-quality annotated training data which is often scarce in many situations. This thesis aims to address this issue by answering the key question: Can limited authentic labeled seismic waveforms be used to generate realistic synthetic ones? To tackle this issue, a generative adversarial network (GAN), a powerful machine learning model known for generating high-quality synthetic data in various domains, is employed.