hewliyang/nisqa-blizzard-challenge-mos
收藏资源简介:
# blizzard-challenge-mos This dataset contains MOS (Naturalness) subjective metrics derived from listening tests done during the evaluation stages of the [Blizzard Challenge](https://www.synsig.org/index.php/Blizzard_Challenge) from 2008 -> 2023. All data is sourced from https://www.cstr.ed.ac.uk/projects/blizzard/data.html. The dataset is formatted for training NISQA models. In particular, at this time there are only two columns for each `.csv` file: - `filepath_deg`: relative file path to the `.wav` file - `mos`: averaged MOS per file over an X number of listeners** **Usage** Similarly, we cannot leverage `datasets` for data loading. Instead, make sure **Git LFS** is installed on your system: ```bash sudo apt install git-lfs git lfs install ``` Download the dataset using: ```bash git clone https://huggingface.co/datasets/hewliyang/blizzard-challenge-mos ``` Untar the wavfiles ```bash find . -name '*.tar.gz' -exec tar -xzf '{}' \; # can remove the tarballs to save disk space rm *.tar.gz ``` **Summary**  Taken from [Mittag. G, Moller. S, 2021](https://arxiv.org/pdf/2104.11673.pdf) To filter for subtasks ie: `EH1`, `EH2` notice that the filenames have been make extremely verbose by design. You can filter easily by checking for substrings ie: `"EH1" in x` **Caveats**** 1. **2014,2015,2016** data only contains system level (ie: submission grouped by participating team). 2. 2015 data was no longer available for download at the link above, hence has been omitted here. **Todos** - [ ] handle 2014 data (??) - [ ] add metadata (license per system, languages, hub/spoke task labels) - [ ] handle 2016 edge cases (listening tests were done on a concatenation of a few raw files + silences. these samples have to be reconstructed which has not been done yet. so we are missing ~500 wavs) - [ ] citations for all Blizzard Challenges **References** ```tex @inproceedings{Mittag_2021, series={interspeech_2021}, title={NISQA: A Deep CNN-Self-Attention Model for Multidimensional Speech Quality Prediction with Crowdsourced Datasets}, url={http://dx.doi.org/10.21437/Interspeech.2021-299}, DOI={10.21437/interspeech.2021-299}, booktitle={Interspeech 2021}, publisher={ISCA}, author={Mittag, Gabriel and Naderi, Babak and Chehadi, Assmaa and Möller, Sebastian}, year={2021}, month=aug, collection={interspeech_2021} } @inproceedings{Mittag_2020, series={interspeech_2020}, title={Deep Learning Based Assessment of Synthetic Speech Naturalness}, url={http://dx.doi.org/10.21437/Interspeech.2020-2382}, DOI={10.21437/interspeech.2020-2382}, booktitle={Interspeech 2020}, publisher={ISCA}, author={Mittag, Gabriel and Möller, Sebastian}, year={2020}, month=oct, collection={interspeech_2020} } ```
# 暴雪挑战赛MOS数据集 本数据集收录了2008年至2023年间,[暴雪挑战赛(Blizzard Challenge)](https://www.synsig.org/index.php/Blizzard_Challenge)各评估阶段通过听音测试得到的MOS(自然度)主观评测指标。所有数据均来源于https://www.cstr.ed.ac.uk/projects/blizzard/data.html。 本数据集的格式专为NISQA模型训练设计。目前每个`.csv`文件仅包含两列: - `filepath_deg`:指向`.wav`音频文件的相对路径 - `mos`:经X名试听者打分后,单文件的平均MOS评分 **使用说明** 此外,本数据集无法通过`datasets`库进行数据加载,请确保你的系统已安装**Git LFS**: bash sudo apt install git-lfs git lfs install 使用以下命令下载数据集: bash git clone https://huggingface.co/datasets/hewliyang/blizzard-challenge-mos 解压音频压缩包: bash find . -name '*.tar.gz' -exec tar -xzf '{}' ; # 可删除压缩包以节省磁盘空间 rm *.tar.gz **数据集概览**  内容取自[Mittag, G, Möller, S, 2021](https://arxiv.org/pdf/2104.11673.pdf) 若需筛选`EH1`、`EH2`等子任务,请注意文件名经过刻意设计,命名格式非常详尽。可通过检索子字符串快速完成筛选,例如:`"EH1" in x` **注意事项** 1. 2014、2015、2016年的数据仅包含系统级结果(即按参赛团队分组的提交结果)。 2. 上述链接已无法下载2015年的数据,因此本数据集未包含该年份的内容。 **待完成事项** - [ ] 处理2014年数据(待确认) - [ ] 添加元数据(各系统的授权协议、语言信息、hub/spoke任务标签) - [ ] 处理2016年的边缘场景:当年的听音测试基于若干原始音频拼接后添加静音段完成,需对这些样本进行重构,目前尚未实现,因此缺失约500个.wav音频文件 - [ ] 补充所有暴雪挑战赛的引用文献 **参考文献** tex @inproceedings{Mittag_2021, series={interspeech_2021}, title={NISQA: A Deep CNN-Self-Attention Model for Multidimensional Speech Quality Prediction with Crowdsourced Datasets}, url={http://dx.doi.org/10.21437/Interspeech.2021-299}, DOI={10.21437/interspeech.2021-299}, booktitle={Interspeech 2021}, publisher={ISCA}, author={Mittag, Gabriel and Naderi, Babak and Chehadi, Assmaa and Möller, Sebastian}, year={2021}, month=aug, collection={interspeech_2021} } @inproceedings{Mittag_2020, series={interspeech_2020}, title={Deep Learning Based Assessment of Synthetic Speech Naturalness}, url={http://dx.doi.org/10.21437/Interspeech.2020-2382}, DOI={10.21437/interspeech.2020-2382}, booktitle={Interspeech 2020}, publisher={ISCA}, author={Mittag, Gabriel and Möller, Sebastian}, year={2020}, month=oct, collection={interspeech_2020} }
数据集概述
数据集名称
- 名称: blizzard-challenge-mos
数据集内容
- 描述: 包含从2008年到2023年Blizzard Challenge评估阶段中获得的MOS(自然度)主观评价指标。
- 数据来源: https://www.cstr.ed.ac.uk/projects/blizzard/data.html
数据格式
- 格式: 每个
.csv文件包含两个列filepath_deg:.wav文件的相对文件路径mos: 每个文件的平均MOS评分,由X名听众评定
数据使用
- 特殊要求: 需要安装Git LFS来处理数据集的下载和解压。
数据限制
- 数据缺失: 2015年的数据由于无法从原链接下载,未包含在本数据集中。
- 数据处理: 2014年和2016年的数据仅包含系统级别的信息,且2016年的某些样本需要重新构建。
未来计划
- 处理2014年数据
- 添加元数据(如许可证、语言、任务标签)
- 处理2016年的边缘案例
- 添加所有Blizzard Challenge的引用
引用信息
- 参考文献:
- Mittag, G., Moller, S., 2021. NISQA: A Deep CNN-Self-Attention Model for Multidimensional Speech Quality Prediction with Crowdsourced Datasets. Interspeech 2021.
- Mittag, G., Moller, S., 2020. Deep Learning Based Assessment of Synthetic Speech Naturalness. Interspeech 2020.




