Datasets:

lighthouse-emnlp2024
/

CASTELLA

language:
  - en
license: cc-by-4.0
size_categories:
  - 1K<n<10K
task_categories:
  - audio-text-to-text
tags:
  - audio-retrieval
  - multimodal
  - moment-retrieval
library_name: lighthouse
configs:
  - config_name: default
    data_files:
      - split: train_en
        path: en/train.json
      - split: train_ja
        path: ja/train.json
      - split: valid_en
        path: en/val.json
      - split: valid_ja
        path: ja/val.json
      - split: test_en
        path: en/test.json
      - split: test_ja
        path: ja/test.json

CASTELLA

This repository provides wav files used in CASTELLA: Long Audio Dataset with Captions and Temporal Boundaries . This dataset is originally provided in GitHub.

Each sample includes long audio containing some audio events with the temporal and textual annotation.

Project page: https://h-munakata.github.io/CASTELLA-demo/

Code: https://github.com/line/lighthouse

This repository only contains annotation data, not audio data. Extracted features are available on HF.

If you need raw audio files, use the download script in GitHub

How to Use

from datasets import load_dataset

dataset = load_dataset("lighthouse-emnlp2024/CASTELLA")

for d in dataset["train_en"]:
    print(d)
    break
>>> {'yid': '-0awng26xQ8', 'global_caption': 'A man is talking while driving, stops somewhere, then drives again', ...}

Citation

@article{munakata2025castella,
  title={CASTELLA: Long Audio Dataset with Captions and Temporal Boundaries},
  author={Munakata, Hokuto and Takehiro, Imamura and Nishimura, Taichi and Komatsu, Tatsuya},
  journal={arXiv preprint arXiv:2511.15131},
  year={2025},
}