AnyTalk introduces a novel framework for generating 3D speech animations for arbitrary characters without necessitating character-specific animation data or manual rigging. The method leverages the motion priors of pre-trained video diffusion models to drive animation. The process begins with Character-specific Fine-tuning (CsF), where a video diffusion model is adapted to a target character using rendered images paired with zeroed-out audio embeddings to represent a "no motion" state. This technique preserves the model's inherent motion capabilities while specializing it to the specific geometry of the character. Once the model generates a talking-head video, the system employs a proposed optimization process to estimate 3D blendshape parameters, effectively uplifting the 2D video into a 3D animation. This approach allows for lip-synced motion across diverse face meshes and various blendshape configurations. To address computational efficiency, the authors developed AnyTalk$_{RT}$, a distilled, streamlined version of the network designed for real-time performance. By utilizing existing video generation models as a foundation, AnyTalk removes the traditional requirement for extensive animation datasets, significantly reducing the manual effort typically required for character rigging and re-meshing in audio-driven animation tasks.