Interactive editing in French Sign Language dedicated to virtual signers: requirements and challenges

Signing avatars are increasingly used as an interface for communication to the deaf community. In recent years, an emerging approach uses captured data to edit and generate sign language (SL) gestures. Thanks to motion editing operations (e.g., concatenation, mixing), this method offers the possibility to compose new utterances, thus facilitating the enrichment of the original corpus, enhancing the natural look of the animation, and promoting the avatar's acceptability. However, designing such an editing system raises many questions. In particular, manipulating existing movements does not guarantee the semantic consistency of the reconstructed actions. A solution is to insert the human operator in a loop for constructing new utterances and to incorporate within the utterance's structure constraints that are derived from linguistic patterns. This article discusses the main requirements for the whole pipeline design of interactive virtual signers, including: (1) the creation of corpora, (2) the needed resources for motion recording, (3) the annotation process as the heart of the SL editing process, (4) the building, indexing, and querying of a motion database, (5) the virtual avatar animation by editing and composing motion segments, and (6) the conception of a dedicated user interface according to user' knowledge and abilities. Each step is illustrated by the authors' recent work and results from the project Sign3D, i.e., an editing system of French Sign Language (LSF) content (http://sign3d.websourd.org/sltat).