ChildTalk: A Multi-Dialect Chinese Child Speech Corpus with Full-Length Child–Caregiver Conversations for Speech Recognition

📖 Overview

ChildTalk is a large-scale, publicly available multi-dialect Chinese child speech dialogue dataset. This dataset solves key problems found in existing Chinese child ASR corpora — mainly their small size, lack of natural conversations, limited dialectal coverage, and missing full-length dialogue recordings. It provides a solid foundation for advancing research in child speech recognition, dialect robustness, dialogue-level modeling, and other related fields. The dataset is released under a non-commercial license, meaning it is available for academic and non-commercial use.

✨ Key Features

• Extensive Dialect Coverage: Features systematic coverage of standard Mandarin, eight Mandarin dialect subgroups (such as Southwestern, Jiao-Liao, and Zhongyuan), and two additional dialects (Southern Min and Jin).
• Conversational Context: Retains full-length sessions that typically last several minutes, making it highly suitable for dialogue-level modeling and context-aware ASR. Experiments show that extending the prompt to previous turns (e.g., $C=3$) provides consistent improvements in models like Whisper and Qwen3-ASR.
• Rich Annotation: Transcriptions faithfully capture child-specific speech patterns, disfluencies, non-speech events (laughter, coughing), and overlapping speech.

📥 Dataset Access

ChildTalk dataset is freely available for non-commercial academic research purposes.

Download

Please download the dataset from ChildTalk.

After that, merge the compressed files and extract them:

cat childtalk.tar.gz.part-* > childtalk.tar.gz
tar -zvxf childtalk.tar.gz

The directory structure after extraction is as follows:

childtalk/
├── information/
│   ├── long/
│   │   ├── all/
│   │   ├── train/
│   │   └── ...
│   ├── short/
│   │   ├── all/
│   │   │   ├── wav.scp
│   │   │   └── text
│   │   ├── train/
│   │   └── ...
│   ├── dialogue_info.csv
│   └── speaker_info.csv
├── long/
│   ├── Beijing/
│   │   ├── U1124_U2124.wav
│   │   └── ...
│   ├── Ji-Lu/
│   └── ...
└── short/
    ├── Beijing/
    │   ├── U1124_U2124_S1_child_7_F_Beijing_light_CHENGDE_Android.wav
    │   └── ...
    ├── Ji-Lu/
    └── ...

• The format of the wav.scp file is .
• The format of the text file is .

⚖️ Ethical Considerations & License

All recordings were collected with prior informed consent from parents or legal guardians, and all data is anonymized. Access to ChildTalk is strictly restricted to non-commercial research use. Users must agree to the data use agreement, which strictly prohibits attempting to infer real-world identities or redistributing the raw audio outside of their research group.

📝 Citation

If you find this corpus useful in your research, please cite our paper:

@inproceedings{zhou-etal-2026-childtalk,
    title = "{C}hild{T}alk: A Multi-Dialect {C}hinese Child Speech Corpus with Full-Length Child{--}Caregiver Conversations for Speech Recognition",
    author = "Zhou, Jiaming  and
      Guo, Yujie  and
      Zhao, Shiwan  and
      Lu, Yao  and
      Wang, Jianye  and
      Sun, Haoqin  and
      Wang, Hui  and
      Qin, Yong",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-acl.251/",
    pages = "5103--5116",
    ISBN = "979-8-89176-395-1"
}