Speechio-Formal

# Speechio-Formal A Chinese ASR evaluation benchmark dataset based on [SpeechIO LEADERBOARD](https://github.com/SpeechColab/Leaderboard), re-annotated with **formal written-style transcriptions**. ## Dataset Description SpeechIO is a widely-used Chinese ASR benchmark covering 27 diverse subsets (ZH00000–ZH00026), spanning domains such as finance, technology, entertainment, news, and daily conversations. This dataset provides: - **`original_text`**: The original verbatim transcription (口语原文，含语气词、重复等) - **`target_text`**: Formal written-style transcription with proper punctuation, cleaned filler words, and normalized expressions (书面语标注，含标点) This annotation is intended for training and evaluating ASR post-correction / text normalization models. ## Dataset Statistics | Split | Subsets | Samples | |-------|---------|---------| | test | 27 | 43,178 | ### Subset Description | Subset ID | Domain | |-----------|--------| | ZH00000 | Finance & Economics | | ZH00001 | Finance & Economics | | ZH00002 | Technology & Internet | | ZH00003 | Technology & Internet | | ZH00004 | Entertainment | | ZH00005 | Entertainment | | ZH00006 | Gaming & Live Streaming | | ZH00007 | Sports Commentary | | ZH00008 | Oral Presentation | | ZH00009 | Oral Presentation | | ZH00010 | News Broadcast | | ZH00011 | News Broadcast | | ZH00012 | TV Drama | | ZH00013 | Variety Show | | ZH00014 | Documentary | | ZH00015 | Audiobook | | ZH00016 | Audiobook | | ZH00017 | Court Recording | | ZH00018 | Conversation | | ZH00019 | Conversation | | ZH00020 | Meeting | | ZH00021 | Meeting | | ZH00022 | Phone Call | | ZH00023 | Phone Call | | ZH00024 | Medical | | ZH00025 | Medical | | ZH00026 | Education | ## Data Fields | Field | Type | Description | |-------|------|-------------| | `idx` | int | Unique sample index | | `audio` | Audio | 16kHz mono WAV audio | | `split` | string | Always `"test"` | | `subset_id` | string | Subset identifier (e.g., `ZH00000`) | | `original_text` | string | Original verbatim transcription | | `language` | string | Always `"Chinese"` | | `target_text` | string | Formal written-style transcription (may be null for some samples) | ## Usage ```python from datasets import load_dataset # Load all subsets dataset = load_dataset("TaurenMountain/Speechio-Formal") # Load a specific subset dataset = load_dataset("TaurenMountain/Speechio-Formal", data_files="data/test-ZH00006.parquet") # Access samples for sample in dataset["train"]: audio = sample["audio"] # {"bytes": ..., "path": ...} original = sample["original_text"] target = sample["target_text"] print(f"Original: {original}") print(f"Target: {target}") ``` ## License This dataset is released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). The audio data originates from the [SpeechIO Leaderboard](https://github.com/SpeechColab/Leaderboard) dataset. Please also comply with the original SpeechIO dataset license terms. ## 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}, } ```