Datasets:

Itbanque
/

ScreenTalk_JA2ZH-XS

task_categories:
  - translation
language:
  - ja
  - zh
tags:
  - translation
  - ja
  - zh_cn
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: 2072186696
      num_examples: 8000
    - name: valid
      num_bytes: 259808873
      num_examples: 1000
    - name: test
      num_bytes: 252154427
      num_examples: 1000
  download_size: 2596980172
  dataset_size: 2584149996
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: valid
        path: data/valid-*
      - split: test
        path: data/test-*

ScreenTalk_JA2ZH-XS

ScreenTalk_JA2ZH-XS is a paired dataset of Japanese speech and Chinese translated text released by DataLabX. It is designed for training and evaluating speech translation (ST) and multilingual speech understanding models. The data consists of spoken dialogue extracted from real-world Japanese movies and TV shows.

📦 Dataset Overview

Source Language: Japanese (Audio)
Target Language: Simplified Chinese (Text)
Number of Samples: 10,000
Total Duration: ~30 hours
Format: Parquet
License: CC BY 4.0
Tasks:
- Speech-to-Text Translation (ST)
- Multilingual ASR+MT joint modeling
- Japanese ASR with Chinese aligned text training

📁 Data Fields

Field Name	Type	Description
`audio`	`Audio`	Raw Japanese speech audio clip
`sentence`	`string`	Corresponding Simplified Chinese text
`duration`	`float`	Duration of the audio in seconds
`uid`	`string`	Unique sample identifier
`group_id`	`string`	Grouping ID (e.g., speaker or scene tag)

🔍 Example Samples

audio	Duration (s)	sentence
JA_00012	4.21	他不会来了。
JA_00038	6.78	为什么你会这样说？告诉我真相。
JA_00104	3.33	安静，有人来了。

💡 Use Cases

This dataset is ideal for:

🎯 Training speech translation models, such as Whisper ST
🧪 Research on multilingual speech understanding
🧠 Developing multimodal AI systems (audio → Chinese text)
🏫 Educational tools for Japanese learners

📥 Loading Example (Hugging Face Datasets)

from datasets import load_dataset

ds = load_dataset("DataLabX/ScreenTalk_JA2ZH-XS", split="train")

📃 Citation

@misc{datalabx2025screentalkja,
  title = {DataLabX/ScreenTalk_JA2ZH-XS: A Speech Translation Dataset of Japanese Audio and Chinese Text},
  author = {DataLabX},
  year = {2025},
  howpublished = {\url{https://huggingface.co/datasets/DataLabX/ScreenTalk_JA2ZH-XS}},
}

We welcome feedback, suggestions, and contributions! 🙌