Reproduce “Gaspard: Expressive unified speech and gesture synthesis with continuous autoregressive flow-matching and decoupled classifier-free guidance” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Téo Guichoux, Théodor Lemerle, Catherine Pelachaud, Nicolas Obin

Gaspard jointly generates speech and gestures autoregressively, allowing each stream’s history to condition the other. It uses decoupled classifier-free guidance to adjust modality-history influence separately, addressing motion collapse from imbalanced feedback. On Seamless-416, its tuned configuration reports kinetic FGD of 0.42 ± 0.33, versus 0.81 ± 0.43 for standard CFG; the guidance settings differ.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.