PD-GS: Phoneme-Driven 3DGS for Audio-Driven Talking Heads

The approach tackles the problem of over-smoothed mouth motion in 3DGS-based talking head rendering. It uses discrete phoneme-level information to better capture brief articulatory events that are often lost in continuous acoustic regression. The method aims to enforce hard constraints such as bilabial closures to reduce artifacts like the 'leaky mouth' effect. This could lead to more realistic and accurate audio-driven avatars for real-time applications.