VOYA Vietnamese Sign Language (VOYA_VSL)

VOYA_VSL is a Vietnamese Sign Language dataset collected and preprocessed for research in sign language recognition.
The dataset consists of sequences of keypoints extracted using MediaPipe, where each sequence represents a sign language video.

📦 Dataset Information

• Format: .npz (NumPy compressed)
• Number of samples: ~161
• Sample structure:
  • sequences: a matrix with shape (60, 1605) (60 frames, 1605 keypoints/features per frame)
  • labels: class index (int32)

📂 Dataset Structure

VOYA_VSL/
│── Merged/
│   ├── class_0001.npz
│   ├── class_0002.npz
│   ├── ...
│   └── class_0161.npz
│── labels.json
│── dataset.py
└── README.md

• Merged/ : contains per-class data stored as .npz files
• labels.json : maps class_xxxx → sign name
• dataset.py : script to load dataset with Hugging Face Datasets
• README.md : dataset description

🛠️ Usage

from datasets import load_dataset

# Load the VOYA_VSL dataset
dataset = load_dataset("Kateht/VOYA_VSL")

print(dataset)

# Example output
DatasetDict({
    train: Dataset({
        features: ['sequences', 'labels'],
        num_rows: ...
    })
})

# Access a sample
sample = dataset["train"][0]
print(sample["sequences"].shape)  # (60, 1605)
print(sample["labels"])           # int label

# Split train/validation/test after loading
dataset = dataset["train"].train_test_split(test_size=0.2, seed=42)

🔖 Labels

The labels.json file maps class IDs to Vietnamese labels. Example:

{
  "class_0001": "địa chỉ (bắc)",
  "class_0002": "địa chỉ (nam)",
  "class_0003": "địa chỉ (trung)"
}

📜 Citation

If you use this dataset in your research, please cite:

@misc{voya_vsl_2025,
  author       = {Kateht et al.},
  title        = {VOYA Vietnamese Sign Language Dataset},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Kateht/VOYA_VSL}}
}