PD-GS uses phoneme tokens to prevent leaky mouth artifacts

Phoneme-Driven Gaussian Splatting (PD-GS) improves audio-driven 3DGS talking heads by integrating time-aligned phoneme tokens from ASR. A Linguistic Fusion Module ensures accurate lip articulation, such as bilabial closures, which are often lost in standard continuous acoustic regression.