Finite Scalar Quantization as Facial Tokenizer for Dyadic Reaction Generation

摘要

Creating a human-like interface in human-robot interaction is a formidable challenge. Many efforts have been made to mimic the human ability of attentive listening and synchronous participation in conversations, especially in terms of facial expressions and head movements. By taking advantage of transformer-based sequence generation models and quantization techniques, this advantage is further enhanced in the areas of text, video, and audio generation. Using Finite Scalar Quantization, we develop a facial expression tokenization module that is able to encode facial expressions in a finite, semantically meaningful vocabulary. Using this module, we establish a more powerful cross-modality transformer-based, non-deterministic model that is able to learn multiple appropriate facial responses in a dyadic conversational context. 1