ESC环境噪音分类数据集

ESC 数据集是一组以统一格式提供的短期环境记录（5 秒长剪辑、44.1 kHz、单通道、Ogg Vorbis 压缩 = 192 kbit/s）。所有剪辑都从通过项目项目获得的公共Freesound.org中提取。有关详细归因列表，请参阅 README 文件。数据集根据知识共享许可证的条款 -归因 - 非商业。 数据集由三部分组成： ESC-US 数据集虽然不是手工注释，但包括原始上传用户提交的标签（标签），这些标签可能用于监督不力的学习（嘈杂和/或缺少标签）。ESC-10 和 ESC-50 数据集已预先安排成 5 个大小均匀的折叠，以便从同一原始源录制中提取的剪辑始终包含在单个折叠中。 标记的数据集也可以作为 GitHub 项目提供：ESC-50 |ESC-10. 如是更详尽的描述和分析，请参阅原始纸张和补充的 IPython 笔记本。 该项目的目标是促进环境声音分类领域的公开研究举措，因为该领域的公开数据集仍然相当稀少。

The ESC dataset is a collection of short-term environmental recordings provided in a unified format: 5-second clips, 44.1 kHz sampling rate, single-channel, compressed with Ogg Vorbis at 192 kbit/s. All clips are extracted from public resources on Freesound.org, which are collected via relevant projects. For a detailed attribution list, please refer to the README file. The dataset is licensed under the Creative Commons Attribution-NonCommercial license.

The dataset consists of three subsets: The ESC-US dataset, which is not manually annotated, but includes tags submitted by the original uploaders. These tags can be utilized for weakly-supervised learning, as they may be noisy and/or have missing labels. Both the ESC-10 and ESC-50 datasets are pre-partitioned into 5 evenly-sized cross-validation folds, ensuring that clips sourced from the same original recording are always assigned to a single fold.

The annotated datasets are also available as GitHub repositories: ESC-50 | ESC-10. For more comprehensive descriptions and analyses, please refer to the original paper and the supplementary Jupyter notebooks.

The overarching goal of this project is to promote open research initiatives in the field of environmental sound classification, as public datasets in this domain remain relatively scarce.

搜集汇总

背景与挑战

背景概述

ESC环境噪音分类数据集包含三个部分：ESC-50、ESC-10和ESC-US，分别提供标记和未标记的环境声音记录，用于环境声音分类研究。数据集格式统一，包含5秒长的音频剪辑，采样率为44.1 kHz，单通道，适用于监督和无监督学习任务。