SICAGE: Speaker-Independent Culture-Aware Gesture Generation using TED4C-L Dataset

The paper tackles a subtle but important problem in co-speech gesture generation: models often look “culture-aware” only because they memorize speaker-specific gesturing habits. SICAGE tries to separate those two factors by learning cultural representations that stay discriminative while being invariant to individual speakers.

The framework is modular. It learns embeddings from audio and text, then uses them to condition a multimodal gesture generator. The authors test two domain generalization strategies for the cultural representation stage: adversarial learning and Fishr regularization. That makes the work more than a dataset drop; it’s also a concrete recipe for reducing speaker leakage in gesture synthesis.

A notable piece for practitioners is ALaDiT, a real-time diffusion-based gesture generator that incorporates the learned cultural embeddings efficiently. The paper claims improvements in motion realism, diversity, beat synchronization, semantic relevance, and cultural consistency, which are exactly the metrics that matter if you’re trying to drive believable NPCs, virtual presenters, or social avatars.

The dataset, TED4C-L, is also a meaningful asset on its own: 106 hours of multimodal data from 764 TED speakers across four cultural groups. For game and interactive media developers, the broader takeaway is that culturally expressive animation is becoming more measurable and less hand-wavy, but only if evaluation avoids speaker-disjoint leakage. That matters for anyone training or benchmarking gesture systems for live agents, cutscene tools, or character...