kalbin/yubao_videos

YuBao: A New Chinese Dialect Speech Benchmark

**Paper** | **Code**

This repository contains the video metadata for the YuBao (語保) dataset, as presented in the paper "Towards Comprehensive Semantic Speech Embeddings for Chinese Dialects".

YuBao is a comprehensive collection from the Chinese Language Resources Protection Project, featuring speech, dialect transcripts, phonetic (IPA) transcriptions, and Mandarin translations for parallel items (1,000 characters, 1,200 words, and 50 sentences) across over 1,300 sites in China.

Dataset Description

The Yubao benchmark is designed to evaluate cross-dialect semantic alignment in speech language models. It covers seven major Chinese dialect subgroups:

• —Mandarin (including dialectal Mandarin)
• —Yue
• —Min (Southern Min/Minnan)
• —Hakka
• —Xiang
• —Wu
• —Gan

This specific dataset (yubao_videos) contains the metadata for video recordings, which is used to facilitate the creation of a speech-to-speech retrieval benchmark.

Data Structure

• —transcript: Dialect transcript
• —translation: Mandarin translation
• —ipa: Phonetic (IPA) transcription
• —english: English translation
• —utterance_id: Unique identifier for the utterance
• —site: Geographic site name
• —subgrouping: Linguistic subgroup classification

To assemble the full audio dataset and run evaluations, please refer to the instructions in the official GitHub repository.

Citation

If you use this benchmark or the associated code, please cite the following:

@article{chang2026yubao,
    author  = {Kalvin Chang and Yiwen Shao and Jiahong Li and Dong Yu},
    title = {Towards Comprehensive Semantic Speech Embeddings for Chinese Dialects},
    journal = {arXiv preprint arXiv: 2601.07274},
    year = {2026}
}

@misc{yubao,
  author       = {{Centre for the Protection of Language Resources of China}},
  title        = {The Chinese Language Resources Protection Project Collection and Display Platform},
  url          = "https://zhongguoyuyan.cn",
  year         = 2023
}

Note: The audio and video copyrights belong to the Centre for the Protection of Language Resources of China.