Ground motion data is essential for both conceptual validation and realistic simulation in civil, structural, and earthquake engineering. However, the availability of recorded waveforms may be limited in regions that lack sufficient seismic instrumentation, and even in well-equipped areas, the data may be unevenly distributed, and this limitation usually obstructs the development of earthquake engineering. This study employs advanced machine learning techniques, specifically variational autoencoder (VAE) and generative adversarial network (GAN). Furthermore, it emphasizes the direct generation of seismic waveforms rather than simulating the focal mechanism and wave propagation, thereby facilitating batch generation. The novel approach, known as VAE-GAN, integrates a GAN in which the generator is structured as a VAE is utilized to synthesize new earthquake waveforms. The generated waveforms indicate that the proposed method can produce a diverse array of waveforms that exhibit reasonable earthquake features, indicating that the artificial earthquake waveforms can still serve as valuable ground motion data. Consequently, the study converged with a discussion and future research.