Research on lip synthesis of virtual digital humans based on the Spark large model

In recent years, with the support of key technologies such as deep learning algorithms, computer vision, natural language processing, and graphics rendering, the production process of virtual digital humans has achieved revolutionary breakthroughs. Lip synthesis technology, as a crucial component, has also garnered widespread attention. Most existing lip synthesis models are trained on English datasets, resulting in poor synthesis effects in Chinese contexts. This paper, based on the study of the Wav2Lip lip synthesis model, constructs a Chinese lip-sync dataset and conducts a series of training fine-tunings to achieve better synthesis effects. By leveraging the multimodal capabilities of the Spark large model, it realizes efficient speech-driven facial animation generation. Furthermore, this paper explores the application potential of the Wav2Lip model, providing a solid theoretical foundation and technical support for future commercial applications.