Do Motion Tokenizers for Co-Speech Gesture Generation Encode Gesture Semantics?

2026-09-28

AI Summary

This study investigates whether discrete motion tokenizers encode semantic gesture attributes in speech-driven co-speech gesture generation. To this end, the authors construct a multi-granularity gesture descriptor framework and employ linear probing to systematically analyze the decodability of properties such as geometry, handedness, and motion categories from codebooks trained via reconstruction objectives. The findings reveal a dissociation between reconstruction quality and semantic decoupling: while geometric features are readily decodable, motion categories remain difficult to recover, demonstrating that relying solely on reconstruction objectives does not guarantee semantic capture. Accordingly, this work proposes a novel evaluation paradigm for motion codebooks grounded in semantic attributes, offering both a theoretical foundation and new perspectives for optimizing discrete motion representations.

Abstract

Discrete motion tokenizers encode motion as atomic units and are widely used for co-speech gesture generation. It remains unclear which motion properties, especially those relevant to gesture semantics, are recoverable from these codebooks. We probe a reconstruction-trained codebook using 19 co-speech gesture descriptors spanning from raw motion to abstract communicative function. Results show that geometry and handedness are readily decodable from token embeddings, while motion category is only weakly decoded despite showing systematic differences in discrete code usage. This gap between reconstruction quality and descriptor decodability suggests that reconstruction objectives alone do not guarantee that gesture semantics are captured, and that evaluating codebooks on such properties can guide the design of more semantic motion tokenizers.

Problem

• co-speech gesture generation
• motion tokenizers
• gesture semantics
• discrete codebook
• reconstruction objective

Innovation

• Motion Tokenizer
• Co-Speech Gesture Generation
• Gesture Semantics
• Codebook Probing
• Discrete Representation