English: ST-News (Shantou News) is a cross-dialect speech-to-subtitle generation dataset derived from broadcast archives of Jinri Shixian (《今日视线》), a news program of Shantou Media Convergence Group. It contains Teochew speech paired with Mandarin news subtitles that follow news subtitle conventions rather than verbatim transcription. This repository hosts the approximately 20-hour public release subset, drawn from 2017 broadcast data.

数据集详情 / Dataset Details

数据集描述 / Dataset Description

中文：ST-News 面向的任务是跨方言语音到字幕生成：给定潮汕话语音，生成符合新闻字幕规范并保留主要新闻命题的普通话文本。由于播音语言为潮汕话而屏幕字幕面向更广泛受众以普通话呈现，语音—字幕对之间存在四类稳定的非逐字映射：词汇规范化、歧义消解、语义重构、命题简化（见术语表）。公开子集包含音频、普通话字幕文本、时间戳和训练/开发/测试清单。

English: ST-News formulates the task of cross-dialect speech-to-subtitle generation: given Teochew speech, the goal is to generate Mandarin text that conforms to news subtitle conventions while preserving the primary news proposition. Because the broadcast speech is in Teochew while the on-screen subtitles are rendered in Mandarin for a broader audience, the speech–subtitle pairs exhibit four types of stable non-verbatim mappings: lexical normalization, ambiguity resolution, semantic reformulation, and propositional simplification (see Glossary). The public subset contains audio, Mandarin subtitle texts, timestamps, and train/validation/test manifests.

策展方 / Curated by: ST-News 项目组（中山大学；北京语言大学；汕头融媒集团）/ ST-News project team (Sun Yat-sen University; Beijing Language and Culture University; Shantou Media Convergence Group)
语言 / Language(s) (NLP): 潮汕话（语音）/ 普通话（字幕文本）；Teochew (speech) / Mandarin Chinese (subtitle text)
许可 / License: ST-News 数据使用协议（见本仓库 LICENSE.md）/ ST-News Data Use Agreement (see LICENSE.md in this repository)

数据集来源 / Dataset Sources

仓库 / Repository: https://huggingface.co/datasets/szcoolj/ST-News
论文 / Paper: Chen, J., Zhang, J., Hu, J., Ke, D., Chen, J., Chen, X., & Huang, W. ST-News: A Shantou News Dataset for Cross-Dialect Speech-to-Subtitle Generation and Non-Verbatim Mapping Analysis. Proceedings of NCMMSC, 2026 (to appear).

用途 / Uses

直接使用 / Direct Use

中文：本数据集仅用于非商业性科学研究，包括：

语音识别与语音到字幕生成模型的训练、评测与基准测试；
跨方言语音理解与非逐字映射分析；
新闻字幕规范及方言—通用语变换研究。

English: The dataset is intended for non-commercial scientific research, including:

Training, evaluation, and benchmarking of speech recognition and speech-to-subtitle generation models;
Cross-dialect speech understanding and non-verbatim mapping analysis;
Research on news subtitle conventions and dialect–standard-language transformation.

禁止用途 / Out-of-Scope Use

中文：以下用途被 ST-News 数据使用协议明确禁止：

任何商业使用（商业诉求请邮件联系授权）；
以任何形式再分发数据集或其任何实质性部分；
创建或发布衍生标注成果；
试图识别、联系或追踪数据集中的任何个人，或任何违反法律法规的使用；
将基于本数据集训练的模型用于商业目的。

此外请注意：字幕并非逐字转写，本数据集不适合用于严格逐字 ASR 结果的评测。

English: The following uses are prohibited by the ST-News Data Use Agreement:

Any commercial use (contact the project team for commercial licensing);
Redistribution of the dataset or any substantial part of it in any form;
Creating or releasing derivative annotations of the dataset;
Attempting to identify, contact, or track any individual appearing in the dataset, or any use violating applicable laws;
Using models trained on the dataset for commercial purposes.

In addition, please note that the subtitles are not verbatim transcripts of the speech; the dataset is not suitable for evaluating strictly verbatim ASR hypotheses against the subtitle side.

数据集结构 / Dataset Structure

中文：每条记录将一段潮汕话节目音频与其普通话字幕文本、句级时间戳和划分归属配对。清单覆盖三个划分：

English: Each record pairs a segment of Teochew broadcast audio with its Mandarin subtitle text, sentence-level timestamps, and split assignment. The manifests cover three splits:

划分 Split	时长 Duration	约合小时 Hours
训练集 Train	18:43:22	≈ 18.72 h
开发集 Validation	00:29:29	≈ 0.49 h
测试集 Test	00:29:03	≈ 0.48 h
合计 Total	19:41:54	≈ 19.70 h

中文：语音覆盖新闻播报、背景旁白、街头采访、专题报道四种场景；说话人包括电视台固定主播与受访市民。

English: Speech covers four scenarios: news broadcasting, background narration, street interviews, and special reports. Speakers include regular TV anchors and interviewed citizens.

数据集创建 / Dataset Creation

策展动机 / Curation Rationale

中文：现有汉语方言语音资源大多采用"方言语音 → 方言正字"的逐字转写框架，未能反映方言语音与通用语言文字之间真实的传播关系。ST-News 来源于真实的新闻生产流程——潮汕话播音与专业制作的普通话字幕——旨在支持跨方言语音到字幕生成研究，以及语音与字幕之间非逐字映射的分析。

English: Existing Chinese dialect speech resources mostly follow a "dialect speech → dialect orthography" verbatim transcription framework, which does not reflect the real-world dissemination relationship between dialect speech and standard-language text. ST-News was created from an actual news production workflow — Teochew broadcast speech with professionally produced Mandarin subtitles — to support research on cross-dialect speech-to-subtitle generation and the analysis of non-verbatim mappings between speech and subtitles.

源数据 / Source Data

数据采集与处理 / Data Collection and Processing

中文：源数据为汕头融媒集团《今日视线》节目播出档案，公开子集取自 2017 年播出数据。原始录像经多媒体解耦流水线处理：音频侧辅助转写与断句、视频侧 OCR 字幕提取、基于时间戳粗匹配加文本相似度校验的自动对齐配对。随后进行跨模态文本一致性检查：一个在潮汕话广播数据上训练的 Conformer 辅助 ASR 模型从音频生成普通话字幕风格假设，计算该假设与 OCR 字幕之间的字符级编辑距离 Δ，剔除 Δ > 3 的低一致性片段。

English: The source data are broadcast archives of Jinri Shixian (《今日视线》), a news program of Shantou Media Convergence Group; the public subset is drawn from 2017 broadcasts. Raw recordings were processed through a multimedia decoupling pipeline: audio-side auxiliary transcription and sentence segmentation, video-side OCR subtitle extraction, and automatic alignment via timestamp-based coarse matching followed by text-similarity verification. A cross-modal text-consistency check was then applied: an auxiliary Conformer-based ASR model trained on Teochew broadcast data produces a Mandarin subtitle-style hypothesis from the audio, and the character-level edit distance Δ between this hypothesis and the OCR subtitle is computed; segments with Δ > 3 were excluded.

源数据生产者 / Who are the source data producers?

中文：语音来自按播音规范播报的电视台固定主播，以及街头采访片段中的受访市民。普通话字幕由节目编辑团队在三审制度（编辑初审—部门主任复审—总编辑终审）下制作完成。

English: The speech is produced by regular TV anchors broadcasting in accordance with broadcasting standards, together with citizens interviewed in street-interview segments. The Mandarin subtitle texts were produced by the program's editorial staff under a three-tier review system (editorial preliminary review — section chief re-review — editor-in-chief final review).

标注 / Annotations

中文：本发布版中的普通话字幕来源于节目真实的编辑制作流程，而非单独的标注过程；公开子集不包含额外标注层。（论文所述四类非逐字映射标签仅人工标注于单独的 358 条评估样本集，不在本发布版中。）

English: The Mandarin subtitles in this release originate from the program's actual editorial workflow rather than from a separate annotation process. No additional annotation layers are included in this public subset. (The four-type non-verbatim mapping labels described in the paper were manually assigned to a separate 358-sample evaluation set and are not part of this release.)

个人与敏感信息 / Personal and Sensitive Information

中文：音频包含可识别个人（新闻主播及受访市民）的声音，新闻内容中可能提及个人姓名，数据未做匿名化处理。依据数据使用协议，使用者不得试图识别、联系或追踪数据集中的任何个人。

English: The audio contains the voices of identifiable individuals (news anchors and interviewed citizens), and news content may mention personal names. No anonymization has been applied. Users must not attempt to identify, contact, or track any individual in the dataset, as required by the Data Use Agreement.

偏差、风险与局限 / Bias, Risks, and Limitations

中文：

领域与方言范围：数据仅限于潮汕话地方新闻节目，在其上训练的模型可能无法泛化到其他领域或方言；
非逐字字幕：字幕侧受新闻字幕规范塑造（存在压缩与改写），不应作为逐字转写的标准答案；
说话人分布倾斜：少数固定主播贡献了大部分音频，市民语音主要为简短的街头采访片段；
时间范围：公开子集为 2017 年播出数据。

English:

Domain and dialect scope: the data are limited to a local news program in Teochew; models trained on it may not generalize to other domains or dialects.
Non-verbatim subtitles: the subtitle side is shaped by news subtitle conventions (compression, rephrasing); it should not be treated as ground-truth verbatim transcription.
Speaker skew: a small number of regular anchors contribute a large share of the audio, while citizen speech occurs mainly in short street-interview clips.
Temporal scope: the public subset covers 2017 broadcasts.

建议 / Recommendations

中文：使用者应将字幕视为字幕风格参考答案而非逐字转写，在报告结果时注意四类非逐字映射的影响，并遵守 LICENSE.md 中的使用限制。

English: Users should treat the subtitles as subtitle-style references rather than verbatim transcripts, report results with awareness of the four non-verbatim mapping types, and comply with the usage restrictions in LICENSE.md.

引用 / Citation

BibTeX：

@inproceedings{chen2026stnews,
  title     = {ST-News: A Shantou News Dataset for Cross-Dialect Speech-to-Subtitle Generation and Non-Verbatim Mapping Analysis},
  author    = {Chen, Jia and Zhang, Jinsong and Hu, Jun and Ke, Dengfeng and Chen, Jian and Chen, Xiaoyun and Huang, Weiqing},
  booktitle = {Proceedings of the National Conference on Man-Machine Speech Communication (NCMMSC)},
  year      = {2026},
  note      = {to appear}
}

APA：

Chen, J., Zhang, J., Hu, J., Ke, D., Chen, J., Chen, X., & Huang, W. (2026). ST-News: A Shantou News Dataset for Cross-Dialect Speech-to-Subtitle Generation and Non-Verbatim Mapping Analysis. Proceedings of the National Conference on Man-Machine Speech Communication (NCMMSC) (to appear). Dataset: https://huggingface.co/datasets/szcoolj/ST-News

术语表 / Glossary

非逐字映射 / Non-verbatim mapping： 语音与字幕之间的稳定偏差，包括词汇替换、结构重组和信息压缩。/ Stable deviations between speech and subtitles, including lexical substitution, structural rephrasing, and information compression.
词汇规范化 / Lexical normalization： 在整体结构基本不变的情况下，将局部方言词汇替换为普通话通用表达。/ Replacing local dialectal lexical items with common Mandarin expressions while the overall structure remains largely unchanged.
歧义消解 / Ambiguity resolution： 结合上下文将直接呈现会造成理解困难的方言成分替换为更清晰的普通话形式。/ Replacing dialectal items that would be difficult to understand if rendered directly, using clearer Mandarin forms based on context.
语义重构 / Semantic reformulation： 超出局部替换的结构重组、语序调整或释义改写。/ Structural reorganization, word-order adjustment, or paraphrasing beyond local substitution.
命题简化 / Propositional simplification： 在保留核心新闻事实的前提下删除和压缩冗余或次要信息。/ Deleting and compressing redundant or secondary information while preserving core news facts.
潮汕话 / Teochew： 闽南语的重要分支，通行于包括汕头在内的潮汕地区。/ A major branch of Southern Min Chinese, spoken in the Chaoshan region including Shantou.

数据集卡片作者 / Dataset Card Authors

ST-News 项目组 / ST-News Project Team

数据集卡片联系方式 / Dataset Card Contact

📧 chenj765@mail2.sysu.edu.cn

Downloads last month: 53

Total file size:

33.2 kB

You need to agree to share your contact information to access this dataset

ST-News 数据集卡片 / Dataset Card for ST-News