Diffusion-Based Co-Speech Gesture Generation Using Joint Text and Audio Representation

Abstract

This paper describes a system developed for the GENEA (Generation and Evaluation of Non-verbal Behaviour for Embodied Agents) Challenge 2023. Our solution builds on an existing diffusion-based motion synthesis model. We propose a contrastive speech and motion pretraining (CSMP) module, which learns a joint embedding for speech and gesture with the aim to learn a semantic coupling between these modalities. The output of the CSMP module is used as a conditioning signal in the diffusion-based gesture synthesis model in order to achieve semantically-aware co-speech gesture generation. Our entry achieved highest human-likeness and highest speech appropriateness rating among the submitted entries. This indicates that our system is a promising approach to achieve human-like co-speech gestures in agents that carry semantic meaning.

State of the Art on Diffusion Models for Visual Computing

“…For example, gesture generation approaches use speech as a condition along with text for style; the attention mechanism is leveraged to synchronise the gestures to the speech. The approach of Deichler et al [DMAB23] proposes a contrastive speech and motion pre‐training module that learns joint embedding of speech and gesture; it learns a semantic coupling between these modalities. DiffuseStyleGesture [YWL*23] is an audio‐driven co‐gesture generation approach that synthesises gestures matching the music rhythm and text descriptions based on cross‐local and self‐attention mechanisms.…”