preprint · arXiv
KoUniTalk: A Lightweight Articulation-Centered Korean-English 3D Talking Face Benchmark
Hyunjung Chung, Unsang Park
In plain English
Korean and English 3D talking-face datasets cannot be trained together because they use different mesh topologies, spatial scales, coordinate systems and sampling rates. The authors retarget VOCASET and released Korean 3D data onto one identity-neutral 1,176-vertex template focused on the mouth region, yielding 22 speakers, 4,978 sequences and 642,781 frames that can be trained jointly without dataset-specific output layers. Their cross-dataset tests show large error gaps when a model trained on one language is tested on the other, which unified training narrows. If you localise avatars beyond English, this is concrete evidence that a meaningful part of the quality gap is a data-format problem rather than a model problem.