Analyzing Talking Head Generation under Real-world Audio Conditions

Audio-driven talking head generation aims to synthesize realistic facial animations that synchronize accurately with the input speech. Although recent models achieve high-quality results under clean and controlled conditions, their robustness to real-world audio variability remains underexplored. We evaluate three representative models using a controlled test set with systematically modified audio input. The experimental results demonstrate that fast speech and increased noise levels degrade synchronization accuracy and visual quality, whereas slower speech improves performance. Both the quantitative and qualitative metrics support these findings. The present study highlights a key limitation of current models—their lack of robustness in unconstrained environments. We conclude by proposing future research directions to improve generalization under real-world conditions.