TED4C-L

TED4C-L is the multimodal dataset introduced with SICAGE: Speaker-Independent Culture-Aware Gesture Generation using TED4C-L Dataset (ECCV 2026). It was created to study co-speech gesture generation across four source groups: Indian, Italian, Japanese, and Turkish TED speakers.

Paper: SICAGE: Speaker-Independent Culture-Aware Gesture Generation using TED4C-L Dataset, accepted at ECCV 2026.

Each record is a five-second window containing aligned motion, speech, and text representations. Windows overlap with a stride of 0.5 seconds. The public release contains derived numeric features, but it does not contain raw video, raw audio, readable subtitles, or transcript strings.

At a Glance

Property	Value
Usable video	106.45 hours
Samples	659,454 overlapping windows
Speakers/source videos	764
Window length	5 seconds
Window stride	0.5 seconds
Motion rate	15 FPS before VQ-VAE encoding
Cultures/source groups	India, Italy, Japan, Turkey
Download size	approximately 189 GiB
Format	LMDB with pickled Python dictionaries

The culture label is a dataset-level source grouping, based on the collection and language organization used to build TED4C-L. It is not a verified label of a person's nationality, ethnicity, identity, or cultural beliefs. It must not be used to infer those attributes for individuals.

Distribution by source group

Label	Samples	Speakers/source videos
Indian	225,575	191
Italian	145,217	194
Japanese	116,056	196
Turkish	172,606	183
Total	659,454	764

Speaker-independent split

The provided split is randomized at the speaker/source-video level, not at the window level. Within each source group, the builder randomly shuffled speaker/video identifiers and assigned approximately 80% to training, 10% to validation, and 10% to testing. It then placed every window from a given identifier in the same subset. The saved release is therefore one fixed randomized split with no speaker/video overlap between subsets.

Split	Samples	Speakers/source videos
Train	531,089	609
Validation	58,523	75
Test	69,842	80

Because adjacent windows overlap heavily, do not reshuffle individual windows across subsets. Doing so can place nearly identical neighboring windows in both training and evaluation data. Use the provided speaker-level split, or create another split by assigning complete speakers/videos rather than individual samples.

What Is in One Sample?

Every LMDB value is a pickled Python dictionary. Arrays are stored in the following layout:

Field	Stored shape	Meaning
motion	(512, 25)	VQ-VAE motion representation; the SICAGE loader returns (25, 512)
audio_mels	(64, 156)	log-mel spectrogram; the loader returns (156, 64)
audio_onsets	(156,)	audio onset strength
audio_wav2vec	(50, 1024)	multilingual Wav2Vec2 speech representation
text_features	(768,)	LaBSE embedding of the aligned source-language subtitle text
text_features_translated	(768,)	LaBSE embedding from an available English subtitle track; otherwise a copy of text_features
translation_source	string	one of translation_disabled, english_subtitles, or empty_text in this release
culture	string	source-group label
speaker	string	source video identifier used as the speaker/domain identifier
language, text_language	string	source and subtitle language metadata
sample_start, sample_end	integer	window boundaries in source-video frame coordinates
motion_fps, scene_fps	float	processed motion and source-scene frame rates

TED4C-L uses LMDB rather than the tabular Hugging Face datasets format, so load it with LMDB or the SICAGE data loader instead of datasets.load_dataset().

Text Features in This Release

The public LMDB retains numeric LaBSE features and provenance only. The readable source and English subtitle strings used to calculate those features were removed before publication. text_features contains the source-language subtitle embedding. text_features_translated uses an available English subtitle track when present; otherwise it copies text_features or stores an empty-text embedding for windows without aligned subtitles.

The SICAGE dataset loader uses text_features by default. Translated features are selected only when use_translated_text=True or the corresponding command-line option is set.

Files

data.mdb
lock.mdb
metadata/
  culture_encodings.pkl
  language_encodings.pkl
  metadata.pkl
  speaker_encodings.pkl
  subject_to_domain_indices.pkl
  training_languages_id_map.pkl
  training_speakers_id_map.pkl
  whole_dataset_splits_subject_independent.pkl

data.mdb contains 659,454 records. The files under metadata/ contain keys, integer encodings, and the official speaker-independent split.

Download

Install huggingface_hub with Xet support, then download the repository snapshot:

python -m pip install -U "huggingface_hub[hf_xet]>=0.36.2,<1.0"
export TED4CL_DATASET_ROOT="/abs/path/to/TED4C-L"

import os
from huggingface_hub import snapshot_download

dataset_root = os.environ["TED4CL_DATASET_ROOT"]
dataset_root = snapshot_download(
    repo_id="ariel-95/TED4C-L",
    repo_type="dataset",
    local_dir=dataset_root,
)
print(dataset_root)

Allow roughly 189 GiB for the download and additional working space for experiments. The main file is very large, so interrupted downloads may take time to recover.

Read the LMDB Directly

Install lmdb, then inspect a record as follows:

import os
import lmdb
import pickle

dataset_root = os.environ["TED4CL_DATASET_ROOT"]
env = lmdb.open(
    dataset_root,
    readonly=True,
    lock=False,
    readahead=False,
    meminit=False,
)

with env.begin(write=False) as transaction:
    cursor = transaction.cursor()
    if cursor.first():
        sample_key = cursor.key().decode("ascii")
        sample = pickle.loads(cursor.value())
        print(sample_key)
        print({name: getattr(value, "shape", type(value).__name__)
               for name, value in sample.items()})

Python pickle must only be loaded from a trusted source. The SICAGE codebase provides the training loader and split utilities in dataset.py. For the repository scripts:

export FULL_DATASET="/path/returned/by/snapshot_download"
export FULL_META="$FULL_DATASET/metadata"

Intended Use

TED4C-L is intended for non-commercial research on:

• co-speech gesture generation and evaluation;
• multimodal speech, text, and motion representation learning;
• reproduction and comparison of the SICAGE experiments; and
• analysis of model behavior across the four dataset source groups.

Do not use TED4C-L for biometric identification, speaker surveillance, nationality or ethnicity inference, decisions about individuals, or claims that the four labels represent cultures in their full diversity.

Limitations and Responsible Use

• TED speakers are public figures in staged talks and are not representative samples of national populations or everyday communication.
• Source-group labels are broad proxies and may encode language, recording, topic, demographic, and production differences in addition to gesture patterns.
• Overlapping windows are highly correlated; sample count is not equivalent to the number of independent observations.
• Motion, speech, and text features may retain information about identity, language, topic, or other personal attributes even though raw media is not included.
• Text embeddings can reflect transcription or alignment errors in the locally obtained source subtitle tracks. This release contains no readable subtitle or transcript strings.
• Models trained on this dataset can reproduce stereotypes or source imbalances. Report results by group and document any additional filtering.

For correction or removal requests, open a discussion in the SICAGE repository.

License and Source Rights

The Hugging Face metadata uses license: other because TED4C-L combines SICAGE-authored dataset organization, metadata, splits, and documentation with numeric representations derived from TED source material. The complete rights notice is in LICENSE.

To the extent the SICAGE authors own the dataset selection, organization, dataset-specific metadata, split definitions, and documentation, those contributions are made available under CC BY-NC 4.0. Users are responsible for making sure their use of any source-derived representation is permitted by the applicable rightsholders.

Citation

If you use TED4C-L or SICAGE, please cite:

@inproceedings{gjaci2026sicage,
  author        = {Gjaci, Ariel and Sgorbissa, Antonio and Murino, Vittorio},
  title         = {SICAGE: Speaker-Independent Culture-Aware Gesture Generation using TED4C-L Dataset},
  booktitle     = {European Conference on Computer Vision (ECCV)},
  year          = {2026},
  eprint        = {2606.30001},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2606.30001},
  url           = {https://arxiv.org/abs/2606.30001}
}