e-gmd
收藏资源简介:
Expanded Groove MIDI Dataset (E-GMD) 是一个大型数据集,包含人类鼓演奏的录音,并配有MIDI标注。该数据集旨在支持鼓转录、音乐信息检索和音频分类等任务,特别是用于提升鼓转录的感知质量。数据内容涵盖来自43个不同鼓套件的音频,总时长为444.5小时。数据集包含45,537个序列,对应1,059个独特的演奏序列,并按照原始Groove MIDI Dataset的划分,分为训练集(35,217个序列,341.4小时)、测试集(5,289个序列,50.9小时)和验证集(5,031个序列,52.2小时)。每个样本包括一个音频文件(WAV格式)和一个配对的MIDI文件,其中MIDI文件提供了鼓演奏的符号化标注。元数据文件记录了音频路径、MIDI路径、原始文件名以及数据划分(split)等信息。数据集以Creative Commons Attribution 4.0 International (CC BY 4.0) 许可证发布,可通过Hugging Face Hub访问,并兼容AudioFolder格式,便于使用Hugging Face datasets库加载。
The Expanded Groove MIDI Dataset (E-GMD) is a large-scale dataset containing recordings of human drum performances with MIDI annotations. It is designed to support tasks such as drum transcription, music information retrieval, and audio classification, particularly for improving the perceptual quality of drum transcription. The dataset covers audio from 43 different drum kits, with a total duration of 444.5 hours. It includes 45,537 sequences corresponding to 1,059 unique performance sequences, and is divided according to the original Groove MIDI Dataset into training set (35,217 sequences, 341.4 hours), test set (5,289 sequences, 50.9 hours), and validation set (5,031 sequences, 52.2 hours). Each sample consists of an audio file (in WAV format) and a paired MIDI file, where the MIDI file provides symbolic annotations of the drum performance. Metadata files record audio paths, MIDI paths, original filenames, and data splits. The dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license, accessible via the Hugging Face Hub, and compatible with the AudioFolder format for easy loading with the Hugging Face datasets library.
数据集名称
Expanded Groove MIDI Dataset (E-GMD)
许可协议
Creative Commons Attribution 4.0 International (CC BY 4.0)
任务类别
音频分类
数据集规模
10,000 < n < 100,000
数据集简介
E-GMD 是一个大型人类鼓点表演数据集,包含带有 MIDI 标注的音频录音。它包含来自 43 个鼓组的 444.5 小时音频,并与原始 Groove MIDI 数据集保持相同的训练/验证/测试集划分。
数据划分
| 划分 | 唯一序列数 | 总序列数 | 时长 |
|---|---|---|---|
| 训练集 | 819 | 35,217 | 341.4 小时 |
| 测试集 | 123 | 5,289 | 50.9 小时 |
| 验证集 | 117 | 5,031 | 52.2 小时 |
| 总计 | 1,059 | 45,537 | 444.5 小时 |
数据存储结构
音频与 MIDI 文件按划分和鼓组目录组织:
audio/{split}/{kit_slug}/{original_drummer/session/path}.wavmidi/{split}/{kit_slug}/{original_drummer/session/path}.midi- 元数据文件:
metadata.csv、metadata/{split}.csv、metadata/all.csv、e-gmd-v1.0.0.csv
元数据字段
元数据文件中包含以下路径列:
file_name:相对于划分文件夹的音频路径audio_path:相对于仓库根目录的音频路径midi_path:相对于仓库根目录的配对 MIDI 路径original_audio_filename:原始存档中的音频路径original_midi_filename:原始存档中的 MIDI 路径
加载方式
可通过 datasets 库加载,支持流式读取:
python
from datasets import load_dataset
ds = load_dataset("schism-audio/e-gmd", split="train", streaming=True)
版本与来源
- 版本:1.0.0
- 原始完整存档 SHA256:
7d9a264fb4c9eabd9fec09d5f8e333192f529b1a1b845d170279a977ac436053 - 官方 Magenta 页面:https://magenta.tensorflow.org/datasets/e-gmd
- Zenodo 记录:https://zenodo.org/records/4300943
引用信息
使用本数据集时,请引用原始 E-GMD 论文并指定版本 1.0.0:
@misc{callender2020improving, title={Improving Perceptual Quality of Drum Transcription with the Expanded Groove MIDI Dataset}, author={Lee Callender and Curtis Hawthorne and Jesse Engel}, year={2020}, eprint={2004.00188}, archivePrefix={arXiv}, primaryClass={cs.SD} }




