SEMANTIX

Semantic Motion Anchors: Bridging Motion and Meaning in Co-Speech Gestures

NeurIPS 2026

SEMANTIX is a human-annotated evaluation set for semantic co-speech gesture understanding. It contains 878 semantic gestures: 778 from 10 TED talks and 100 from the BEAT2 motion-capture dataset.

Each gesture comes with the speech it accompanies and an expert description of two things:

• Physical form: handedness, hand shape, orientation, position and trajectory.
• Communicative intent: what the gesture conveys in context.

Use it to evaluate gesture description and text–gesture retrieval with our code.

Example (ted_1010, from Wendy Suzuki's talk):

Field	Value
Transcript	to be better in me.
Gesture words	in me
Physical form	Both hands, with index fingers and thumbs extended while other fingers are closed, point inward toward the body at chest level.
Communicative intent	refer to self experience
Intent category	Self-reference
Watch	On TED's YouTube channel, at 5:41

Dataset configurations

Configuration	Examples	Contents
all (default)	878	Fields shared by both sources
ted	778	TED annotations, timestamps and YouTube links that open at each gesture
beat2	100	BEAT2 annotations, frame indices and avatar videos
ted_talks	10	Speaker, title, ted.com page and YouTube upload of each TED talk
beat2_anchors	17,106	Generated anchors of all BEAT2 training, validation and test gestures in the paper (see Add-on)

The four SEMANTIX configurations have a single test split. The beat2_anchors add-on loads as one train split that contains all three experimental splits; filter by its split column.

All SEMANTIX rows have id, transcript, gesture_words (the words spoken during the gesture), gold_description, start_sec and end_sec. On top of that:

• TED rows add physical_form, communicative_intent, one of 15 intent_category labels, youtube_url/youtube_embed_url, and link_check (see Videos).
• BEAT2 rows add the BEAT2 sequence, frame indices and an avatar video_file. Their description is a single sentence that ends with the intent.

Load

pip install datasets huggingface_hub

from datasets import load_dataset

semantix = load_dataset("aboots/SEMANTIX", split="test")
ted = load_dataset("aboots/SEMANTIX", "ted", split="test")
beat2 = load_dataset("aboots/SEMANTIX", "beat2", split="test")

example = ted[0]
print(example["transcript"], "|", example["gesture_words"])
print(example["physical_form"], "->", example["communicative_intent"])

Videos

TED. SEMANTIX was annotated on local copies of the TED talks. TED's usage policy does not allow sharing clips of its talks, so we do not host TED video. Each row links to the same talk on TED's official YouTube channel instead:

• youtube_url opens the talk at the gesture.
• youtube_embed_url requests playback covering the gesture interval, rounded outwards to whole seconds:

<iframe width="560" height="315" src="https://www.youtube.com/embed/BHY0FxzoKZE?start=341&end=344" allowfullscreen></iframe>

The annotations and timestamps are unchanged. We compared the YouTube uploads with our local copies: they have the same length, and their frames align across each talk.

We then checked every link with word-level speech recognition of the talks:

• 751 of the 778 rows: speech recognition detects every content word of the gesture within 0.5 s of the linked interval.
• 17 more: within 2 s.
• 3 rows: the words are 2.2–2.4 s away.
• 2 rows: the words are 5–8 s away; these are flagged.
• 5 rows: recognition missed a word (a spelling or recognition variant), and TED's own transcript has it at the linked time.

Each row's result is in the link_check column. VALIDATION.md lists word timestamps for every row outside the interval. start_sec and end_sec hold the exact times.

BEAT2. Each row links to a 640×480, 30 fps render of the motion on an SMPL-X avatar, with the speaker's audio. The render covers the gesture with up to 5 s of context on each side.

from huggingface_hub import hf_hub_download

video = hf_hub_download("aboots/SEMANTIX", beat2[0]["video_file"], repo_type="dataset")

For the 3D motion, slice the BEAT2 SMPL-X sequence with the end-exclusive, 30 fps frame indices:

import numpy as np

row = beat2[0]
seq = np.load(f"beat_english_v2.0.0/smplxflame_30/{row['beat2_sequence']}.npz", allow_pickle=True)
poses = seq["poses"][row["start_frame"]:row["end_frame"]]

Sources and data notes

• TED: the gestures come from the annotated gesture regions of Ram et al. (UIST 2025) in 10 TED Expressive talks. TED transcripts are Whisper ASR output.
• BEAT2: the gestures come from BEAT2 (Liu et al., CVPR 2024).
• Annotation: descriptions follow the gesture annotation schema of Kipp (2005).
• Review: a second expert reviewed the first 231 TED annotations. For those rows, reviewed_description holds the reviewed text, which is the reference used in the paper's anchor-quality evaluation.
• Shared form text: within a gesture-lexicon entry (gesture_lexicon_id), TED clips share the same physical-form text.
• Known issues are marked in the flags column, and annotation text is otherwise released as written.

VALIDATION.md lists the flags, text clean-up, checksums and link checks.

Add-on: semantic motion anchors

SEMANTIX is the human-annotated set above. The addons/ folder holds the inputs of the retrieval experiments in the paper: machine-generated anchors for BEAT2 and the TED evaluation texts. It is not part of the gold annotations and is only needed to reproduce those experiments with our code.

File	Rows	Contents
addons/beat2_anchors.csv (config beat2_anchors)	17,106	Generated semantic motion anchor of every BEAT2 English deictic, iconic and metaphoric gesture, its physical-form (physical_form) and communicative-intent (communicative_intent) parts, an LLM-assigned intent category (intent_category, one of the 15 SEMANTIX categories, not human-verified), the clip's transcript window and its train/val/test split in the paper
addons/beat2_motion_narratives.jsonl	17,106	Verbalized motion narrative of each gesture: the body and hand description of every 8-frame motion token, which the anchor-generation prompt receives with the transcript
addons/ted_eval.csv	778	The SEMANTIX TED gestures with the transcript query (query_text) and the human-written gesture-lexicon description (lexicon_description, from Ram et al.) used to represent TED gestures in the paper's TED and cross-dataset evaluations

anchors = load_dataset("aboots/SEMANTIX", "beat2_anchors", split="train")
train_anchors = anchors.filter(lambda row: row["split"] == "train")

• Generation: anchors were generated with GPT-5.4 from the motion narrative and the transcript, and split into physical form and intent with Qwen3-8B. Two anchors could not be split and have empty parts. Four anchors contain a stray ;" left by a spreadsheet export; they are kept exactly as used for training.
• Clips: clip matches the clip files written by the code's BEAT2 preparation script (<sequence>---g<index>---<type>---<label>).
• TED queries: query_text equals gesture_words except for three rows (ted_0450, ted_0804, ted_1422), where the evaluation used the original region text.
• License: the add-on is released under CC BY-NC 4.0 like the annotations; the BEAT2 transcripts it contains come from BEAT2.

License

• Annotations: SEMANTIX annotations are released under CC BY-NC 4.0 for non-commercial research.
• TED talks: owned by TED (CC BY-NC-ND 4.0). SEMANTIX links to TED's official uploads and does not redistribute them.
• BEAT2: the BEAT2 gestures, avatar videos and audio come from BEAT2. If you use the BEAT2 portion, please also cite BEAT2 (see below).

Citation

If you use this dataset, please cite our paper:

@misc{suresh2026semanticmotionanchorsbridging,
      title={Semantic Motion Anchors: Bridging Motion and Meaning in Co-Speech Gestures},
      author={Varsha Suresh and Mohammad Mahdi Abootorabi and Mohamed Salman and M. Hamza Mughal and Christian Theobalt and Ashwin Ram and Jürgen Steimle and Vera Demberg},
      year={2026},
      eprint={2605.30608},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.30608},
}

If you use the BEAT2 portion, please also cite BEAT2:

@inproceedings{liu2024emage,
  title     = {EMAGE: Towards Unified Holistic Co-Speech Gesture Generation via Expressive Masked Audio Gesture Modeling},
  author    = {Liu, Haiyang and Zhu, Zihao and Becherini, Giorgio and Peng, Yichen and Su, Mingyang and Zhou, You and Zhe, Xuefei and Iwamoto, Naoya and Zheng, Bo and Black, Michael J.},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages     = {1144--1154},
  year      = {2024}
}

Questions? Open an issue on GitHub.