“Sign-to-Speech Prosody Transfer via Sign Reconstruction-based GAN” has been ACCEPTED for ICPR2026. （T.Manabe, M1）

This work introduces the “Sign-to-Speech Prosody Transfer” task, which directly integrates sign language prosody — including emphasis and intonation — into synthesized speech without text mediation. To preserve prosodic nuances lost in conventional two-stage pipelines (sign language → text → speech), we propose S2PFormer, a transformer-based model leveraging sign language prosody reconstruction that enables training on unpaired datasets.

Authors: Toranosuke Manabe, Yuto Shibata, Shinnosuke Takamichi, Yoshimitsu Aoki

Project: Coming soon