ayj111/openstbench-paired-set

OpenSTBench LibriTTS-based Paired Speaker Set

This dataset contains the LibriTTS-based paired speaker set constructed for speaker preservation evaluation in OpenSTBench, a multidimensional benchmark for speech translation systems.

It is designed to support evaluation of whether speech-to-speech translation systems preserve speaker characteristics when generating translated speech.

• —Paper: https://arxiv.org/abs/2605.30792
• —HF Paper page: https://huggingface.co/papers/2605.30792
• —Code: https://github.com/sjtuayj/OpenSTBench

Dataset Description

The dataset provides paired speech samples and metadata for evaluating speaker preservation in speech-to-speech translation. Each example includes:

• —an original LibriTTS utterance,
• —a speaker prompt audio sample,
• —a synthesized translated speech sample,
• —source and target text metadata,
• —duration information,
• —optional quality-related metadata such as COMET-KIWI scores.

The dataset is intended for evaluation and research use in speech translation, speaker similarity, and speech generation assessment.

Usage

from datasets import load_dataset

dataset = load_dataset("ayj111/openstbench-paired-set")

sample = dataset["train"][0]

print(sample["id"])
print(sample["english_text"])
print(sample["target_text"])

original_audio = sample["original_audio"]
prompt_audio = sample["prompt_audio"]
synthesized_audio = sample["synthesized_audio"]

print(original_audio)
print(prompt_audio)
print(synthesized_audio)

Dataset Structure

Each sample contains the following fields:

Intended Use

This dataset is intended for:

• —evaluating speaker preservation in speech-to-speech translation systems,
• —comparing speaker similarity metrics,
• —analyzing the relationship between translation quality and speaker preservation,
• —supporting reproducibility of OpenSTBench experiments.

It is not intended as a general-purpose speaker cloning dataset or as a training dataset for impersonation systems.

License and Terms

The OpenSTBench metadata is released as part of the OpenSTBench project. The audio data is derived from LibriTTS and synthesized speech samples. Users are responsible for complying with the license and usage terms of the original data sources and any synthesis components involved.

Because the dataset combines metadata, source audio, and synthesized speech, the license is marked as other. Please refer to the original data providers and the OpenSTBench repository for details.

Citation

If you use this dataset or OpenSTBench in your research, please cite:

@misc{an2026openstbenchsemanticevaluationspeech,
      title={OpenSTBench: Beyond Semantic Evaluation for Speech Translation},
      author={Yanjie An and Yuxiang Zhao and Yichi Zhang and Qixi Zheng and Yujie Tu and Keqi Deng and Kai Yu and Xie Chen},
      year={2026},
      eprint={2605.30792},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2605.30792},
}

Contact

For questions about OpenSTBench, please refer to the GitHub repository:

https://github.com/sjtuayj/OpenSTBench