WenetSpeech-Formal

# WenetSpeech-Formal ## Dataset Description **WenetSpeech-Formal** is a Chinese speech dataset built upon [WenetSpeech](https://github.com/wenet-e2e/wenetspeech), augmented with formal written-style transcriptions. Each audio sample contains two levels of transcription: | Field | Description | |-------|-------------| | `original_text` | Original verbatim transcription（口语原文，含口语习惯表达） | | `target_text` | Formal written-style transcription（正式书面体转写，由大语言模型辅助生成） | This dataset is designed for training models that convert spoken Chinese to formal written Chinese, such as ASR post-processing, spoken-to-written normalization, and speech understanding. ## Dataset Statistics | Split | # Samples | Notes | |-------|-----------|-------| | train | 969,202 | WenetSpeech L set | | validation | 13,562 | WenetSpeech DEV set | | test | 31,933 | WenetSpeech TEST_NET + TEST_MEETING | ## Data Fields - **`audio`**: 16kHz mono WAV audio clip - **`original_text`**: Verbatim transcription of the spoken utterance (口语原文) - **`target_text`**: Formal written-style transcription — the core new annotation of this dataset - **`split`**: Source split identifier (e.g., `train`, `validation`, `test-net`, `test-meeting`) - **`language`**: `Chinese` ## Example ```python { "audio": {"array": [...], "sampling_rate": 16000}, "original_text": "但是我想这里这当中就是如果如果一定要那个挑一点儿什么的话", "target_text": " 但是，如果一定要从中挑出一点什么的话。", "split": "test-meeting", "language": "Chinese" } ``` ## Usage ```python from datasets import load_dataset # Load the full dataset ds = load_dataset("TaurenMountain/WenetSpeech-Formal") # Access a training sample sample = ds["train"][0] print(sample["original_text"]) # spoken text print(sample["target_text"]) # formal written text audio_array = sample["audio"]["array"] # numpy array, 16kHz ``` ## Source Data This dataset is derived from [WenetSpeech](https://huggingface.co/datasets/speechcolab/wenetspeech) (CC-BY 4.0). The `target_text` formal annotations are newly created for this release. ## License [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) ## Citation If you use this dataset, please cite our work: ```bibtex @misc{ning2026formalasrendtoendspokenchinese, title={FormalASR: End-to-End Spoken Chinese to Formal Text}, author={Wanyi Ning and Yinshang Guo and Haitao Qian and Jiyuan Cheng and Wei Zhou and Weiyuan Feng and Yufei Zhang}, year={2026}, eprint={2605.19266}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2605.19266}, } ```