SemTalk: Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis

Abstract

Co-speech gesture generation must balance frequent rhythm-aligned movements with sparse but essential semantic gestures. We present SemTalk, a holistic co-speech gesture generation framework with frame-level semantic emphasis. Our key insight is to separately learn general motions and sparse motions, and then adaptively fuse them. In particular, rhythmic consistency learning is explored to establish rhythm-related base motion, ensuring a coherent foundation that synchronizes gestures with the speech rhythm. Subsequently, semantic emphasis learning is designed to generate semantic-aware sparse motion, focusing on frame-level semantic cues. Finally, to integrate sparse motion into the base motion and generate semantic-emphasized co-speech gestures, we further leverage a learned semantic score for adaptive synthesis. Qualitative and quantitative comparisons on two public datasets demonstrate that our method outperforms the state-of-the-art, delivering high-quality co-speech motion with enhanced semantic richness over a stable base motion. The code is available at https://github.com/Xiangyue-Zhang/SemTalk.

Demo

Method

SemTalk uses a two-stream design: one stream learns a stable rhythm-aligned base, while another stream activates sparse semantic motion only when speech calls for emphasis.

Semantic Emphasis

The central design choice is not to make every frame more expressive. Instead, SemTalk learns when the speech contains a phrase that should trigger stronger, sparse gestures.

Qualitative Results

The visual comparisons show what the semantic branch changes: gestures become more intentional at meaningful words while staying stable during ordinary speech.

Quantitative Results

Lower is better for FGD, MSE, and LVD. BC and DIV follow the benchmark convention shown in each table. SemTalk now provides results for the original Speaker 2 protocol and a released 25-speaker BEAT2 checkpoint.

BibTeX

@inproceedings{zhang2025semtalk,
  title={SemTalk: Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis},
  author={Zhang, Xiangyue and Li, Jianfang and Zhang, Jiaxu and Dang, Ziqiang and Ren, Jianqiang and Bo, Liefeng and Tu, Zhigang},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={13761--13771},
  year={2025}
}

Positioning and Related Work

SemTalk: frame-level semantic emphasis in co-speech gesture generation.

SemTalk learns rhythm-related base motion and semantically expressive sparse motion separately. A learned frame-level semantic score controls their fusion. This makes the rhythm and semantic components explicit in the generation process. The focus is when speech calls for a meaningful gesture beyond rhythmic motion. See the SemTalk paper for the method and original experiments.

Related methods address different parts of co-speech gesture generation. EMAGE (paper) jointly learns audio-to-gesture generation and masked gesture reconstruction for holistic motion. SemGes (paper) uses semantic coherence and relevance learning to connect gestures with speech semantics. GestureLSM (paper) models interactions between tokenized body regions and uses latent shortcut learning for flow-matching sampling. SemTalk focuses on the separation and frame-level fusion of rhythmic and semantic motion. These distinctions describe modeling objectives, without ranking results across different evaluation protocols.

BEAT2 baseline use. Keep the original Speaker 2 paper protocol separate from the later 25-speaker release. Their results are not a controlled single-speaker versus multi-speaker ablation. The all-speaker workflow and released metrics and checkpoint metadata document that release's setup.