Teochew-News-20h：来自2017年播出数据的约20小时语音—字幕子集，实际时长约19.70小时，全部属于 Δ=0 数据，包含音频、普通话节目字幕、时间戳以及训练／验证／测试划分清单。该子集是792小时 TeochewSpeech-News 语料的一部分，不是额外的独立语料。
Teochew-Eval-Sample：完整5小时 Teochew-Eval 的公开样本，共304条语句（约15.1分钟），在新闻播报、背景旁白、街头采访三类场景间近似均衡分布。每条语句包含同步视频、音频、经人工核验的节目字幕、关键帧图像及元数据标注。

完整792小时 TeochewSpeech-News 语料和完整5小时 Teochew-Eval 均不随本仓库公开。

English: Teochew-Open is a public resource package for research on Teochew television-news speech, derived from broadcast archives of Jinri Shixian (《今日视线》), a news program of Shantou Media Convergence Group. The repository contains two distinct public components:

Teochew-News-20h: an approximately 20-hour speech–subtitle subset drawn from 2017 recordings, with an actual duration of approximately 19.70 hours. All samples belong to the Δ=0 stratum. The release includes audio, Mandarin program subtitles, timestamps, and train/validation/test manifests. This subset is part of the 792-hour TeochewSpeech-News corpus, not an additional independent corpus.
Teochew-Eval-Sample: a public sample of the complete five-hour Teochew-Eval set, containing 304 utterances (approximately 15.1 minutes) distributed approximately evenly across news broadcasting, background narration, and street interviews. Each utterance includes synchronized video, audio, a manually verified program subtitle, a keyframe image, and metadata annotations.

Neither the complete 792-hour TeochewSpeech-News corpus nor the complete five-hour Teochew-Eval set is publicly released in this repository.

数据集详情 / Dataset Details

数据集描述 / Dataset Description

中文：本仓库同时公开新闻语音训练／研究数据和四模态评测样本，但两者的内容、质量口径与复现范围不同：

Teochew-News-20h 提供约19.70小时的潮汕话语音及对应的普通话节目字幕。公开清单包含约18.72小时训练集、0.49小时验证集和0.48小时测试集，可用于数据检查、模型训练与验证，以及相关研究的数据划分复现。
Teochew-Eval-Sample 用于检视 Teochew-Eval 的数据格式和开展小样本评测研究。论文主ASR结果在完整5小时 Teochew-Eval 上计算，不能仅用这304条公开样本复现。
Teochew-Eval 按播出日期和节目期次从792小时语料中留出，不参与训练或ASR模型更新。训练与评测分区之间存在说话人重叠，因此 Teochew-Eval 不是说话人独立评测集。
两个组件的文本均为普通话节目字幕。Teochew-Eval 的参考字幕经过人工核验，但不是独立制作的逐字转写；基于该参考文本的CER属于字幕参考口径下的ASR评价。

Δ 表示辅助ASR输出与OCR字幕窗口中最佳连续片段之间的最小字符级编辑距离。Δ=0 仅表示两路自动文本存在精确匹配片段，不表示字幕必然完全正确，也不等同于经过人工逐字核验。Teochew-News-20h 的质量口径为 Δ=0；Teochew-Eval-Sample 则同时满足 Δ=0 且节目字幕经过人工核验。

English: The repository provides both television-news speech for training and research and a four-modality evaluation sample. These components differ in content, quality criteria, and reproduction scope:

Teochew-News-20h provides approximately 19.70 hours of Teochew speech with corresponding Mandarin program subtitles. Its manifests define approximately 18.72 hours for training, 0.49 hours for validation, and 0.48 hours for testing. It supports data inspection, model training and validation, and reproduction of the data partitions used in related research.
Teochew-Eval-Sample supports inspection of the Teochew-Eval data format and small-sample evaluation studies. The main ASR results in the paper were computed on the complete five-hour Teochew-Eval set and cannot be reproduced from the 304 public utterances alone.
Teochew-Eval is held out from the 792-hour corpus by broadcast date and program episode and is used neither for training nor for ASR updating. Speakers overlap between the training and evaluation partitions; Teochew-Eval is therefore not a speaker-independent benchmark.
Both components use Mandarin program subtitles as text references. Teochew-Eval subtitles are manually verified, but they are not independently produced verbatim transcripts. CER against these references therefore follows a subtitle-referenced ASR protocol.

Δ is the minimum character-level edit distance between an auxiliary ASR output and its best-aligned contiguous fragment within an OCR subtitle window. Δ=0 denotes an exact match between the two automatic text streams; it does not guarantee that the subtitle is completely correct or has been manually verified verbatim. Teochew-News-20h uses the Δ=0 criterion, whereas Teochew-Eval-Sample additionally requires manual verification of the program subtitles.

策展方 / Curated by: TeochewSpeech-News 项目组（中山大学；北京语言大学；汕头融媒集团；深圳广播电影电视集团）/ TeochewSpeech-News project team (Sun Yat-sen University; Beijing Language and Culture University; Shantou Media Convergence Group; Shenzhen Media Group)
语言 / Language(s): 潮汕话（语音）／普通话（节目字幕文本）；Teochew (speech) / Mandarin Chinese (program subtitle text)
许可 / License: Teochew-News-20h 采用 ST-News 数据使用协议；Teochew-Eval-Sample 采用 Teochew-Eval 数据使用协议。适用条款见本仓库 LICENSE.md 及随附协议。/ Teochew-News-20h is governed by the ST-News Data Use Agreement, and Teochew-Eval-Sample is governed by the Teochew-Eval Data Use Agreement. See LICENSE.md and the accompanying agreements in this repository.

数据集来源 / Dataset Sources

仓库 / Repository: https://huggingface.co/datasets/szcoolj/Teochew-Open
主要论文 / Primary paper: Chen, J., Ma, Y., Hu, J., Ke, D., Chen, J., Chen, X., Huang, W., Chen, Y., Lin, H., He, X., & Chen, Z. TeochewSpeech-News: Construction and Recognition of a Teochew Dialect Literary-Reading Speech Corpus with a Human-Seeded Multimodal Framework. ACM TALLIP, 2026 (under revision).
相关论文 / Related 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. NCMMSC 2026 (to appear).

用途 / Uses

直接使用 / Direct Use

中文：本数据集面向非商业性科学研究，两个组件的推荐用途分别为：

Teochew-News-20h

潮汕话及低资源方言语音识别模型的训练、验证与测试；
潮汕话语音到普通话新闻字幕的生成与对齐研究；
数据格式、时间戳和训练／验证／测试划分的检查与复现；
相关ST-News等时长跨语料比较的数据侧复现。该复现还需使用 Teochew-Wild 原始划分及相同模型与训练配置，本仓库不包含 Teochew-Wild。

Teochew-Eval-Sample

Teochew-Eval 数据格式、四模态对齐方式和元数据字段的检查；
小样本ASR评测、视听语音识别（AVSR）与多模态对齐研究；
按新闻播报、背景旁白和街头采访分场景开展分析。

English: The dataset is intended for non-commercial scientific research. Recommended uses differ by component:

Teochew-News-20h

Training, validation, and testing of Teochew and other low-resource dialect ASR models;
Generation and alignment of Mandarin news subtitles from Teochew speech;
Inspection and reproduction of data formats, timestamps, and train/validation/test partitions;
Data-side reproduction of the matched-duration cross-corpus comparison reported in the related ST-News study. Such reproduction additionally requires the original Teochew-Wild splits and the same model and training configuration; Teochew-Wild is not included in this repository.

Teochew-Eval-Sample

Inspection of the Teochew-Eval data format, four-modality alignment, and metadata fields;
Small-sample ASR evaluation, audio-visual speech recognition (AVSR), and multimodal alignment research;
Scene-specific analysis of news broadcasting, background narration, and street interviews.

禁止用途与使用边界 / Out-of-Scope Use and Usage Boundaries

中文：使用者须分别遵守两个组件适用的数据使用协议。主要限制包括：

未经许可不得进行商业使用；商业诉求须联系项目组取得授权；
不得以任何形式再分发数据或其任何实质性部分；
不得试图识别、联系或追踪数据中的任何个人，或从事任何违反法律法规的使用；
Teochew-Eval-Sample 的衍生标注发布、模型训练及商业模型使用限制，以 Teochew-Eval 数据使用协议为准；
Teochew-News-20h 仅限科研用途，其具体限制以 ST-News 数据使用协议为准。

Teochew-Eval-Sample 属于评测基准的一部分。将其用于训练会影响后续评测的可比性，建议仅用于评测和数据检查。无论使用哪个组件，都不得将 Δ=0 直接解释为人工核验或无误转写。

English: Users must comply with the data use agreement applicable to each component. The principal restrictions include:

No commercial use without authorization; contact the project team regarding commercial licensing;
No redistribution of the data or any substantial part of it in any form;
No attempt to identify, contact, or track any individual appearing in the data, and no use that violates applicable laws;
Restrictions on derivative annotations, training use, and commercial use of models for Teochew-Eval-Sample are governed by the Teochew-Eval Data Use Agreement;
Teochew-News-20h is released for research use only, subject to the ST-News Data Use Agreement.

Teochew-Eval-Sample is part of an evaluation benchmark. Using it for training affects the comparability of subsequent evaluations and is therefore discouraged. For either component, Δ=0 must not be interpreted as manual verification or an error-free transcript.

数据集结构 / Dataset Structure

中文：仓库按两个组件组织。以下为逻辑结构；Teochew-News-20h 的具体文件名、音频格式和清单字段以仓库实际文件为准。

English: The repository is organized into two components. The following is a logical overview; exact filenames, audio formats, and manifest fields for Teochew-News-20h follow the files actually provided in the repository.

Teochew-Open/
├── Teochew-News-20h/       # 音频、普通话节目字幕、时间戳及划分清单
│                           # audio, Mandarin program subtitles, timestamps, and manifests
└── Teochew-Eval-Sample/    # 304条四模态评测样本 / 304 four-modality evaluation samples

Teochew-News-20h

公开内容 / Released content：

内容 Content	说明 Description
音频 / Audio	与字幕和时间戳对应的潮汕话语音 / Teochew speech aligned with subtitles and timestamps
节目字幕 / Program subtitles	普通话新闻节目字幕，不应默认视为独立逐字转写 / Mandarin program subtitles, not independent verbatim transcripts by default
时间戳 / Timestamps	用于语音与字幕对齐的时间信息 / Temporal information for speech–subtitle alignment
划分清单 / Split manifests	训练、验证、测试三个分区 / Train, validation, and test partitions

公开划分 / Public splits：

分区 Split	时长 Duration
Train 训练集	≈ 18.72 h
Validation 验证集	≈ 0.49 h
Test 测试集	≈ 0.48 h
合计 Total	≈ 19.70 h

注：分项时长经四舍五入，合计按公开数据总时长报告。/ Note: split durations are rounded, while the total follows the reported duration of the public release.

Teochew-Eval-Sample

每条语句对应一个四模态数据包，目录按场景组织：/ Each utterance corresponds to a four-modality bundle, organized by scene:

Teochew-Eval-Sample/
├── News/                  # 新闻播报 / news broadcasting
│   └── <group_id>/        # 数据组（同一节目片段）/ data group (same program segment)
│       ├── <uid>.mp4      # 视频片段 / video clip
│       ├── <uid>.mp3      # 音频流 / audio stream
│       ├── <uid>.txt      # 经人工核验的普通话节目字幕 / manually verified Mandarin program subtitle
│       └── <uid>.jpg      # 关键帧图像 / keyframe image
├── Bkgd/                  # 背景旁白 / background narration
│   └── ...
└── Street/                # 街头采访 / street interviews
    └── ...

元数据字段 / Metadata fields：

字段 Field	含义 Meaning	示例 Example
`uid`	语句唯一标识 / utterance ID	`News1`
`scene`	场景标签 / scene label	`News` / `Bkgd` / `Street`
`transcript`	经人工核验的节目字幕 / manually verified program subtitle	`海豚的生长环境并不乐观`
`gender`	说话人性别 / speaker gender	`M` / `F`
`age_group`	年龄段 / age group	`Youth` / `Mid.` / …
`role`	说话人角色 / speaker role	`Anchor` / `Actor` / `Citizen`
`duration_s`	时长（秒）/ duration (seconds)	`3.18`

公开样本统计 / Public sample statistics：

场景 Scene	语句数 Utterances	时长 Duration
News 新闻播报	101	≈ 4.6 min
Bkgd. 背景旁白	102	≈ 4.8 min
Street 街头采访	101	≈ 5.7 min
合计 Total	304	≈ 15.1 min

数据集创建 / Dataset Creation

策展动机 / Curation Rationale

中文：Teochew-Open 将新闻语音—字幕数据与四模态评测样本集中发布，以支持低资源方言语料构建、识别评价和数据格式检查。Teochew-News-20h 提供带公开划分的新闻语音研究数据；Teochew-Eval-Sample 提供可检视的多模态评测实例。两个组件分别发布，避免将20小时训练／研究子集与304条评测样本混为同一数据分区或同一复现口径。

English: Teochew-Open brings together television-news speech–subtitle data and a four-modality evaluation sample to support low-resource dialect corpus construction, recognition evaluation, and data-format inspection. Teochew-News-20h provides news speech with public splits for training and research, while Teochew-Eval-Sample provides inspectable multimodal evaluation instances. Keeping the two components distinct prevents the 20-hour research subset and the 304 evaluation utterances from being conflated in data partitioning or reproduction claims.

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2015—2025年播出档案。原始录像经人工播种的 DiaVision-Frame 多模态流程处理，包括字幕区域检测、OCR字幕提取、辅助ASR、跨模态时间同步、基于 Δ 的文本一致性分层，以及依托潮汕话发音词典和声学模型的强制对齐。该流程分别刻画文本一致性与声学可对齐性，不将任何一路自动输出直接视为无误转写。

Teochew-News-20h 来自2017年数据的 Δ=0 分区。Teochew-Eval 按播出日期和节目期次从792小时语料中留出，其节目字幕经过人工核验。评测样本的场景标签和说话人属性由视听融合方法生成。

English: The source data are 2015–2025 broadcast archives of Jinri Shixian (《今日视线》), a news program of Shantou Media Convergence Group. Recordings are processed through the human-seeded DiaVision-Frame multimodal workflow, including subtitle-region detection, OCR subtitle extraction, auxiliary ASR, cross-modal temporal synchronization, Δ-based textual-agreement stratification, and forced alignment supported by a Teochew pronunciation lexicon and acoustic model. Textual agreement and acoustic alignability are treated as separate checks, and neither automatic output is directly assumed to be an error-free transcript.

Teochew-News-20h is drawn from the Δ=0 partition of the 2017 recordings. Teochew-Eval is held out from the 792-hour corpus by broadcast date and program episode, and its program subtitles are manually verified. Scene labels and speaker attributes in the evaluation samples are generated through audiovisual fusion.

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

中文：语音来自电视台主播及节目中的受访市民。普通话节目字幕由节目编辑团队制作。公开数据涉及的地方新闻生产过程具有稳定的字幕与播报规范，但采访语音仍包含自发口语和复杂声学条件。

English: The speech is produced by television-news anchors and citizens interviewed in the programs. Mandarin program subtitles are produced by the program's editorial staff. The local-news production process follows established subtitle and broadcasting conventions, while interview speech may contain spontaneous language and challenging acoustic conditions.

标注 / Annotations

中文：Teochew-News-20h 提供对齐的普通话节目字幕、时间戳和划分清单，不声称全部字幕经过人工逐字核验。Teochew-Eval-Sample 的节目字幕经过人工核验，但仍属于节目字幕参考文本，而非独立制作的逐字转写。场景标签与说话人属性由视听融合方法生成。训练与评测分区按播出日期和节目期次隔离，但不保证说话人独立。

English: Teochew-News-20h provides aligned Mandarin program subtitles, timestamps, and split manifests; it does not claim that every subtitle has been manually verified verbatim. Teochew-Eval-Sample subtitles are manually verified but remain program-subtitle references rather than independently produced verbatim transcripts. Scene labels and speaker attributes are generated through audiovisual fusion. Training and evaluation partitions are separated by broadcast date and program episode, but speaker independence is not guaranteed.

个人与敏感信息 / Personal and Sensitive Information

中文：公开音频包含可识别个人的声音，评测样本还包含新闻主播及受访市民的影像，新闻内容可能提及个人姓名。数据未做匿名化处理。依据适用的数据使用协议，使用者不得试图识别、联系或追踪数据中的任何个人。

English: The public audio contains identifiable voices, and the evaluation sample additionally contains images of news anchors and interviewed citizens. News content may mention personal names. No anonymization has been applied. Users must not attempt to identify, contact, or track any individual in the data, as required by the applicable Data Use Agreement.

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

中文：

领域与来源范围：数据来源于地方电视新闻节目，结论可能无法直接泛化到其他领域、媒体类型或方言。
公开子集代表性：Teochew-News-20h 仅来自2017年且全部为 Δ=0 数据，不能代表完整792小时语料的年份、场景和一致性分层分布。
字幕文本性质：文本为遵循新闻制作规范的普通话节目字幕，可能与潮汕话语音存在编辑、规范化或省略；CER应按字幕参考口径解释。
Δ=0 的能力边界：精确OCR—ASR匹配仍可能包含共同错误、截断或其他字幕差异，不等于人工核验或完全正确。
说话人分布与重叠：主播和受访市民的语音分布不均，训练与评测分区之间存在说话人重叠。
评测样本规模：304条约15.1分钟样本适合数据检查和小样本研究，不足以复现论文基于完整5小时 Teochew-Eval 的主ASR结果。
场景差异：新闻播报、背景旁白和街头采访的声学条件及语言风格不同，街头采访通常包含更多环境噪声与自发口语现象。

English:

Domain and source scope: the data are derived from local television-news programs, and findings may not directly generalize to other domains, media types, or dialects.
Representativeness of the public subset: Teochew-News-20h contains only 2017 recordings from the Δ=0 stratum and does not represent the year, scene, or agreement-stratum distribution of the complete 792-hour corpus.
Nature of the subtitle text: the text consists of Mandarin program subtitles shaped by news-production conventions and may contain editing, normalization, or omission relative to the Teochew speech. CER should therefore be interpreted under the subtitle-referenced protocol.
Limits of Δ=0**:** an exact OCR–ASR match may still contain shared errors, truncation, or other subtitle discrepancies and does not imply manual verification or complete correctness.
Speaker distribution and overlap: speech is unevenly distributed across anchors and interviewed citizens, and speakers overlap between training and evaluation partitions.
Evaluation-sample scale: the 304 utterances (approximately 15.1 minutes) support data inspection and small-sample studies but are insufficient to reproduce the paper's main ASR results computed on the complete five-hour Teochew-Eval set.
Scene differences: news broadcasting, background narration, and street interviews differ in acoustic conditions and language style, with street interviews generally containing more environmental noise and spontaneous speech.

建议 / Recommendations

中文：

明确报告所使用的组件、版本和数据分区，不将 Teochew-News-20h 的测试分区与 Teochew-Eval 混用。
报告CER时注明文本归一化方式，并说明评价采用节目字幕参考文本。
使用 Teochew-Eval-Sample 时按场景分别报告结果，不将其结果直接与完整5小时评测集上的论文主表混同。
复现相关等时长跨语料比较时，使用本仓库公开清单、Teochew-Wild 原始划分及相同训练配置，并引用ST-News论文。
遵守 LICENSE.md 及各组件随附协议中的使用限制。

English:

Clearly report the component, version, and data partition used; do not conflate the Teochew-News-20h test partition with Teochew-Eval.
Specify text normalization when reporting CER and identify program subtitles as the reference text.
Report scene-specific results for Teochew-Eval-Sample and do not conflate them with the paper's main tables computed on the complete five-hour evaluation set.
To reproduce the related matched-duration cross-corpus comparison, use the public manifests in this repository, the original Teochew-Wild splits, and the same training configuration, and cite the ST-News paper.
Comply with LICENSE.md and the agreement accompanying each component.

引用 / Citation

主要论文 / Primary paper:

@article{chen2026teochewspeechnews,
  title   = {TeochewSpeech-News: Construction and Recognition of a Teochew Dialect Literary-Reading Speech Corpus with a Human-Seeded Multimodal Framework},
  author  = {Chen, Jia and Ma, Yilin and Hu, Jun and Ke, Dengfeng and Chen, Jian and Chen, Xiaoyun and Huang, Weiqing and Chen, Yingcong and Lin, Hongwei and He, Xia and Chen, Zhanghua},
  journal = {ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP)},
  year    = {2026},
  note    = {under revision}
}

使用 Teochew-News-20h 或复现相关等时长比较时，请同时引用 / When using Teochew-News-20h or reproducing the related matched-duration comparison, please also cite:

@inproceedings{chen26stnews,
  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 2026 National Conference on Man-Machine Speech Communication (NCMMSC)},
  year      = {2026},
  note      = {To appear}
}

APA：

Chen, J., Ma, Y., Hu, J., Ke, D., Chen, J., Chen, X., Huang, W., Chen, Y., Lin, H., He, X., & Chen, Z. (2026). TeochewSpeech-News: Construction and Recognition of a Teochew Dialect Literary-Reading Speech Corpus with a Human-Seeded Multimodal Framework. ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP) (under revision). Dataset: https://huggingface.co/datasets/szcoolj/Teochew-Open

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 2026 National Conference on Man-Machine Speech Communication (NCMMSC) (to appear).

术语表 / Glossary

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.
文白异读 / Wen-Bai variation： 同一汉字在文读与白读中对应不同音系形式的现象；在本研究语境中主要用于说明发音词典和强制对齐资源的建设背景。/ The phenomenon in which the same character maps to distinct phonological forms in literary and colloquial readings; in this work, it primarily informs the pronunciation resources used for forced alignment.
Δ：辅助ASR输出与OCR字幕窗口中最佳连续片段之间的最小字符级编辑距离。/ The minimum character-level edit distance between an auxiliary ASR output and its best-aligned contiguous OCR fragment.
Δ=0： OCR与ASR片段精确匹配，不等于人工核验或完全无误。/ An exact OCR–ASR fragment match, which does not imply manual verification or complete correctness.
四模态数据包 / Four-modality bundle： Teochew-Eval-Sample 中每条语句保留的同步视频、音频、节目字幕与关键帧图像。/ The synchronized video, audio, program subtitle, and keyframe image retained for each Teochew-Eval-Sample utterance.
场景标签 / Scene labels： 新闻播报（News）、背景旁白（Bkgd）、街头采访（Street）三类语音场景。/ Three speech scenarios: news broadcasting (News), background narration (Bkgd), and street interviews (Street).
DiaVision-Frame： 用于构建本数据资源的人工播种多模态处理流程，分别结合文本一致性检查与方言特定的强制对齐。/ The human-seeded multimodal workflow used to construct these resources, combining textual-agreement checks with dialect-specific forced alignment.

数据集卡片作者 / Dataset Card Authors

TeochewSpeech-News 项目组 / TeochewSpeech-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:

2.1 GB

You need to agree to share your Hugging Face information to access this dataset

Teochew-Open 数据集卡片 / Dataset Card for Teochew-Open