CSRTE

CSRTE is a dataset for Chinese speech relation extraction. Each sample is a Mandarin speech clip with a transcript, entity spans, and relation triplets. The task is to extract (head, relation, tail) triplets directly from the audio.

• 19,855 annotated clips, 33.78 hours, 16 kHz mono
• 10 entity types and 49 relation types
• Paper: ACRE (Findings of EMNLP 2026)
• Code: MWXGOD/CSRTE

Splits

Split	Clips	Hours	Entities	Relations
train	15,000	25.52	37,743	18,439
validation	2,000	3.42	5,000	2,465
test	2,855	4.84	6,921	3,345
total	19,855	33.78	49,664	24,249

validation corresponds to dev.json in the original release.

Fields

Field	Description
audio	16 kHz mono audio (FLAC, 16-bit)
sentence	Chinese transcript
triplets	Relation triplets as {head, relation, tail}, e.g. {庹庆明, 居住于, 荥经}
entities	Entity spans: id, start, end (character offsets, end exclusive), text, label
relations	Raw relation annotations: from_id, to_id, labels, direction
audio_id	Clip ID, e.g. common_voice_zh-CN_18776086 or BAC009S0185W0305
source	Common Voice 17.0 or AISHELL-1
speaker_id	AISHELL-1 speaker ID (e.g. S0185). Empty for Common Voice clips, which have no speaker ID
duration_seconds	Clip length in seconds

triplets is built from relations and entities. head is the text of the from_id entity, tail is the text of the to_id entity, and relation is the middle part of the label (人物:居住于:地点 → 居住于). The full label with entity types is still in relations.

The entity types are <人物>, <地点>, <组织>, <称号>, <时间>, <金钱>, <事件>, <著作>, <产品> and <规则条例>. The full list of relation types is in annotations/dataset_info.json.

Example:

{
  "sentence": "庹庆明，四川荥经人，汉族，中国民主同盟盟员。",
  "triplets": [
    {"head": "庹庆明", "relation": "居住于", "tail": "荥经"},
    {"head": "荥经", "relation": "位于", "tail": "四川"}
  ],
  "entities": [
    {"id": "Uh6LiuB7Am", "start": 0, "end": 3, "text": "庹庆明", "label": "<人物>"},
    {"id": "P9qwmYXbpx", "start": 6, "end": 8, "text": "荥经", "label": "<地点>"}
  ]
}

Usage

from datasets import load_dataset

ds = load_dataset("DMU-ITREC/CSRTE")            # train / validation / test
sample = ds["train"][0]
print(sample["sentence"], sample["triplets"])
waveform = sample["audio"]["array"]              # 16 kHz

To browse without downloading everything, pass streaming=True.

Other files

• annotations/train.json, dev.json, test.json: the original annotation files, including the data_info speaker metadata (gender, age, accent, etc.)
• release_manifest.json: file checksums and row counts

Audio processing

• Common Voice clips: we converted the 16 kHz float64 NumPy arrays to 16-bit PCM. 125 clips peaked above 1.0. We scaled each of them down by its peak to avoid clipping.
• AISHELL-1 clips: we re-encoded the 16-bit WAV files to FLAC without changing any samples.
• We did no resampling, trimming, or denoising.

Known issues

• In five training samples, an entity's text does not match the corresponding slice of sentence. These are BAC009S0230W0457, BAC009S0346W0368, BAC009S0710W0321, BAC009S0718W0319 and BAC009S0907W0481. We kept them as they are.
• 715 clips have no entities and 9,321 have no relations. Their lists are empty.
• Speaker metadata is incomplete, and the splits are not speaker-disjoint.

Source data and license

Source	Clips	License
Common Voice 17.0 (zh-CN)	9,908	CC0 1.0
AISHELL-1	9,947	Apache 2.0

The CSRTE annotations are released under CC BY 4.0. The audio stays under the license of its source corpus. Do not try to identify the speakers.

Citation

@inproceedings{mu-etal-2026-acre,
  title = "{ACRE}: An Acoustic-Cue and Contextual-Knowledge Reasoning Enhancement Framework for Speech Relation Extraction",
  author = "Mu, Wenxuan and Ning, Jinzhong and Zhao, Di and Zhang, Yijia",
  booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2026",
  year = "2026",
  publisher = "Association for Computational Linguistics",
  pages = "26343--26352",
  url = "https://aclanthology.org/2026.findings-emnlp.1244/",
  doi = "10.18653/2026.findings-emnlp.1244"
}

Questions or corrections: please open a discussion.