NN-Based Czech Sign Language Synthesis

• Published in: Speech and Computer Vol. 11658; pp. 559 - 568
• Main Authors: Zelinka, Jan, Kanis, Jakub, Salajka, Petr
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2019; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: 3D skeleton reconstruction; Implicit translation; Sign language synthesis
• ISBN: 3030260607; 9783030260606
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-26061-3_57

This paper describes our Czech sign language synthesis that converts a Czech text into a series of skeletal poses. Our main goal is to avoid demanding handcrafted annotations of videos and to avoid a manual mapping between sign language glosses and skeletal poses. Thus, instead of solving these task separately, we join a model of an implicit neural-network-based translator and a model of the mapping between sign language glosses and we train both models together. For this purpose, we propose a simple differentiable operation that decomposes input symbols and it allows to produce a required series without any recurrent mechanism. We used The OpenPose toolbox to automatically extract skeletal poses and we designed a gradient-descend-based algorithm that converts a 2D skeleton model to a 3D skeleton model in order to fix misplaced and missing joints. Weather forecast parts of The daily news in Czech sign language were used to obtain our training and testing data. Our experiments demonstrate the benefit of the implicit translator and an ability of the designed sign language synthesis system to produce naturally formed skeletal poses.