ViSL-News

Dataset Summary

ViSL-News is a sentence-level Vietnamese Sign Language (VSL) dataset constructed from sign-interpreted Vietnamese news broadcasts.

The dataset was built from HTV Tin Tức videos published on YouTube during 2024–2025. Each sample consists of a sentence-level sign-language video clip paired with a Vietnamese text sentence.

ViSL-News was constructed using ViSL-Tool, a semi-automated framework designed for news videos that contain spoken Vietnamese and an on-screen Vietnamese Sign Language interpreter but do not necessarily provide subtitles.

The framework combines:

• video collection and filtering,
• automatic speech recognition (ASR),
• sentence segmentation and temporal alignment,
• text normalization,
• sentence-level clip generation,
• signer grouping,
• and expert verification for quality control and evaluation-set curation.

ViSL-News is intended to support research on:

• Vietnamese Sign Language Translation,
• gloss-free sign language translation,
• sentence-level video-to-text learning,
• weakly supervised dataset construction,
• and data-centric research for sign language processing.

Languages

• Sign language: Vietnamese Sign Language (VSL)
• Text language: Vietnamese

Dataset Structure

Each sample corresponds to one sentence-level video-text pair.

Depending on the released version, fields may include:

• sample_id
• video
• text
• start_time
• end_time
• signer_id
• split
• source_url

The exact field names may differ across released versions.

Dataset Creation

Source Data

ViSL-News was constructed from HTV Tin Tức broadcasts published on YouTube during 2024–2025 that contain Vietnamese Sign Language interpretation.

The source videos contain:

• spoken Vietnamese audio,
• an on-screen sign-language interpreter,
• and naturally occurring broadcast-news content.

The dataset focuses on the news domain.

Creation Pipeline

ViSL-News was created using the ViSL-Tool pipeline.

The main stages are:

1. collect candidate news videos,
2. select a predefined signer region of interest (ROI),
3. detect shot boundaries and filter unsuitable videos,
4. detect speech regions,
5. transcribe spoken Vietnamese using ASR,
6. segment transcripts into sentences,
7. perform forced alignment to obtain sentence timestamps,
8. normalize transcription errors,
9. cut sentence-level video clips,
10. extract face embeddings and group samples by signer,
11. perform expert review on selected samples,
12. export the final dataset.

The automatic pipeline is designed to reduce the amount of manual annotation required while retaining expert review for quality control.

Annotation Process

The dataset combines automatically generated annotations with expert verification.

Automatic labels are produced using:

• speech recognition,
• temporal alignment,
• sentence segmentation,
• text normalization,
• and signer clustering.

Expert verification was applied to samples selected for the validation and test sets.

For each reviewed sample, experts could:

• edit the transcript,
• adjust start and end timestamps,
• correct signer metadata when needed,
• approve the sample,
• or reject unsuitable samples.

A two-pass expert verification protocol was used for the evaluation data.

Dataset Statistics

ViSL-News contains:

• 26,395 sentence-level samples
• approximately 67 hours of signing video
• 9,918-word vocabulary
• 5 signers
• video resolution of 326 × 426
• 25 FPS

Sentence statistics:

• minimum sentence length: 7 words
• maximum sentence length: 53 words
• mean sentence length: 29.43 words

Clip-duration statistics:

• minimum duration: 2 seconds
• maximum duration: 21 seconds
• mean duration: 9.14 seconds

Signer distribution:

Signer	Samples
Signer 1	10,307
Signer 2	6,624
Signer 3	5,234
Signer 4	2,471
Signer 5	1,759
Total	26,395

Dataset Splits

The released benchmark uses the following split:

Split	Samples	Share	Vocabulary	Unseen words	Signers
Train	24,399	92.44%	9,918	-	5
Validation	997	3.78%	1,969	0	5
Test	999	3.78%	1,987	0	5
Total	26,395	100%	9,918	-	5

The validation and test sets were selected so that their words also occur in the training set.

All three splits contain samples from all five signers. Therefore, the current benchmark is signer-dependent and does not evaluate generalization to unseen signers.

Quality Control

During dataset construction:

• 9,411 source videos were processed,
• 3,096 videos were retained after filtering,
• filtering precision was 99.52%,
• filtering retention rate was 32.90%,
• filtering noise rate was 0.48%.

Automatic annotation initially produced 26,527 sentence-level samples.

During signer clustering and review:

• 94 embeddings were marked as noise,
• 5 main signer clusters were identified,
• 2 additional small clusters were flagged for inspection,
• 132 unsuitable samples were removed,
• resulting in the final 26,395 samples.

For evaluation-set verification:

• 2,000 samples were reviewed,
• 1,996 samples were retained,
• 4 samples were rejected.

Two experts participated in the verification process.

Benchmark Results

Two gloss-free Sign Language Translation baselines were evaluated on ViSL-News.

Model	BLEU-1	BLEU-2	BLEU-3	BLEU-4	ROUGE-L
SpaMo	16.47	11.83	7.56	5.25	16.05
GFSLT-VLP	13.66	8.41	4.69	3.63	12.31

These results are intended as initial reference points rather than a comprehensive model comparison.

Intended Uses

ViSL-News is intended for:

• academic research,
• educational use,
• non-commercial scientific research,
• Vietnamese Sign Language Translation,
• gloss-free SLT,
• sentence-level video-to-text modeling,
• weakly supervised learning,
• temporal alignment research,
• dataset construction and curation research.

Out-of-Scope Uses

The dataset is not intended for:

• commercial deployment,
• commercial redistribution,
• monetized redistribution of source media,
• surveillance or identity recognition,
• or uses that violate the rights of the original content owners.

The signer identifiers included in the dataset are intended as dataset-level grouping metadata and should not be interpreted as real-world identity labels.

Copyright and Usage Notice

ViSL-News was constructed for non-commercial research from publicly accessible sign-interpreted HTV Tin Tức broadcasts published on YouTube.

The maintainers do not claim ownership of the original broadcast videos or audio.

The rights to the original source content remain with their respective rights holders.

The original source videos and audio are not redistributed as part of this dataset release.

The release should contain only research-related data such as:

• annotations,
• metadata,
• sentence-level information,
• and source URLs,

subject to the terms of the released version.

Users are responsible for complying with the terms of the original source platforms and applicable copyright law.

If a legitimate rights holder requests removal or modification of affected material or metadata, the maintainers will review the request and cooperate where appropriate.

Limitations

ViSL-News has several important limitations:

• The dataset is restricted to the broadcast-news domain.
• It does not represent all Vietnamese Sign Language usage contexts.
• The source material contains interpreted news signing rather than unconstrained everyday conversations.
• Sentence boundaries and timestamps are initially produced automatically and may contain alignment noise.
• Training labels are weakly supervised by spoken-language transcripts.
• The dataset contains only five signers, resulting in limited signer diversity.
• The signer distribution is naturally imbalanced.
• The current train/validation/test split is signer-dependent.
• Therefore, the current benchmark does not measure generalization to unseen signers.
• Automatic temporal alignment may introduce noise, particularly for long or complex sentences.

Ethical Considerations

The dataset is intended for non-commercial research.

Signer metadata is used only to group samples for dataset analysis and experimental splitting. The system does not attempt to infer the real-world identity of signers.

Researchers should avoid using this dataset for biometric identification, surveillance, or other purposes unrelated to sign-language research.

Citation

If you use ViSL-News or ViSL-Tool, please cite the corresponding paper, thesis, or official dataset release.

Paper

@article{visl_tool_2026,
  title   = {ViSL-Tool: A Semi-Automated Framework for Constructing Sentence-Level Vietnamese Sign Language Datasets from News Videos},
  author  = {Anonymous Author(s)},
  year    = {2026},
  note    = {Manuscript}
}