metadata
dataset_info:
features:
- name: audio
dtype: audio
- name: duration
dtype: float64
- name: sentence
dtype: string
- name: uid
dtype: string
- name: group_id
dtype: string
splits:
- name: train
num_bytes: 38952870042.272
num_examples: 146157
- name: valid
num_bytes: 4613208092.88
num_examples: 18270
- name: test
num_bytes: 4372566889.24
num_examples: 18270
download_size: 47749925884
dataset_size: 47938645024.392
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: valid
path: data/valid-*
- split: test
path: data/test-*
π ScreenTalk_JA2ZH
ScreenTalk_JA2ZH is a bilingual speech dataset containing Japanese audio paired with Chinese (Simplified) sentence translations. It is specifically curated from film and TV content, making it highly suitable for tasks such as speech translation, multilingual ASR fine-tuning, and movie subtitle generation.
π Dataset Summary
- Audio Language: Japanese (ja)
- Text Language: Chinese (Simplified, zh-CN)
- Domain: Film and television dialogues
- Format: Each sample contains a path to a Japanese audio clip and its corresponding Chinese sentence translation.
π Dataset Structure
Each sample in the dataset has the following fields:
Field | Description |
---|---|
audio |
Path or link to the Japanese audio file |
sentence |
Chinese translation of the spoken content |
Example:
{
"audio": "clips/0001.wav",
"sentence": "ζ们εΎεΏ«ηΉη¦»εΌθΏιγ"
}
π‘ Use Cases
- Fine-tuning ASR models (e.g., Whisper) for Japanese input and Chinese output
- Training speech translation models
- Movie subtitle generation in bilingual format
- Evaluation for multilingual speech understanding systems
π οΈ How to Load
from datasets import load_dataset
dataset = load_dataset("DataLabX/ScreenTalk_JA2ZH")
π License
CC BY-NC 4.0 This dataset is made available for research and non-commercial use only.
π§ Citation
If you use this dataset, please cite:
@misc{datalabx2025screentalkja2zh,
title = {ScreenTalk_JA2ZH: Japanese Audio with Chinese Translations},
author = {DataLabX},
year = {2025},
howpublished = {\url{https://huggingface.co/datasets/DataLabX/ScreenTalk_JA2ZH}},
}