Generating animations of photorealistic 3D head avatars is important for many applications like digital humans, movie production, and immersive telepresence. To advance this task, Gaussian Splatting, a state-of-the-art neural scene representation technique, was used in recent papers. Gaussian Splatting offers key advantages such as high-fidelity rendering and real-time performance (high FPS), making it an ideal tool for creating high-quality avatars. A key challenge in audio-driven Gaussian avatars is their reliance on person-specific training, requiring multi-view images of one single person. This severely limits scalability and generalization ability. In the project, we aim to improve the generalization ability of audio-driven Gaussian avatars, enabling them to generate high-quality animations for unseen individuals using only a single image or sparse-view images as input.