SynStard-1000

Dataset Summary

SynStard-1000 is a 1,000-hour synthetic dataset for training and evaluating end-to-end speech-to-speech translation (S2ST) models. It is built from English-Chinese parallel texts in the WMT News Commentary v18 corpus and contains approximately 390,000 sentence pairs with paired synthetic speech.

Dataset Structure

.
├── map/
│   └── all.tsv
│── text/
│   ├── en/
│   │   ├── en.txt
│   │   ├── en_1.txt
│   │   ├── ...
│   │   ├── en_11.txt
│   │   ├── en_extra.txt
│   ├── zh/
│   │   ├── zh_cleaned.txt
│   │   ├── zh_cleaned_1.txt
│   │   ├── ...
│   │   ├── zh_cleaned_11.txt
│   │   ├── zh_cleaned_extra.txt
└── wavs/
│   ├── en_1/
│   │   ├── 4501/
│   │   │   ├── 00004501-00000000-en.wav
│   │   │   ├── ...
│   │   ├── 4504/
│   │   ├── ...
│   ├── ...
│   ├── zh_1/
│   │   ├── 4501/
│   │   │   ├── 00004501-00000000-zh.wav
│   │   │   ├── ...
│   │   ├── 4504/
│   │   ├── ...
│   ├── ...

Data Format

Each row in all.tsv corresponds to a speech translation pair with associated audio paths and transcriptions. :

zh_path	en_path	zh_text	en_text
wavs/zh_1/4501/00004501-00000000-zh.wav	wavs/en_1/4501/00004501-00000000-en.wav	因此我认为那些相信会实现强劲复苏的预测者是过于乐观了，最终实际的经济发展也与我的预测一致。	Thus, I thought that forecasters of a stronger recovery were looking at the world through rose-tinted glasses. Economic developments unfolded much as I anticipated.

License

This dataset is licensed under the apache-2.0.

Citation

If you use SynStard-1000 in your work, please cite:

@inproceedings{pu2025empowering,
  title={Empowering Large Language Models for End-to-End Speech Translation Leveraging Synthetic Data},
  author={Pu, Yu and Liu, Xiaoqian and Zhang, Guangyu and Yan, Zheng and Zhang, Wei-Qiang and Chen, Xie},
  booktitle={Proc. Interspeech 2025},
  pages={26--30},
  year={2025}
}