An audio dataset for sound effect variations

This is a sound effect dataset constructed to test neural network's capability of generating diverse categories of sounds. It is composed of two categories: Footstep-set and Impact-set. All sounds are sourced from Freesound. The Footstep-set contains 9 categories with 2127 sounds in total while Impact-set has 7 categories with 1666 sounds in total. Each category is a different variation of the main dataset. For example, the Footstep-set contains footsteps recorded on metal boards and on concrete floors, each showing slightly different timbres of sounds. The Impact-set contains impulsive sounds commonly used in games such as gunshots, explosion, punch, etc. 

Each dataset is separated into training (90%) and testing (10%). All sounds are sampled in 44.1khz, 16bit, 4-second mono files. The metadata for each dataset records the index of the sound, category, ID for retrieving the link of Freesound, Sound Creator, Name for the sound file, and the train/test set split.

This dataset was used to train and test our model in a recent paper that was submitted but under review. We will upload the accompanying paper once it's been accepted.

本数据集为音效数据集，旨在测试神经网络生成多样化类别声音的能力。本数据集包含两个类别：脚步声数据集（Footstep-set）与冲击音效数据集（Impact-set），所有声音素材均来源于Freesound平台。
脚步声数据集包含9个子类别，总计2127条音频；冲击音效数据集则包含7个子类别，总计1666条音频。每个子类别均为主数据集下的不同变体。例如，脚步声数据集包含在金属板与水泥地面录制的脚步声，二者呈现出略有差异的音色特征；冲击音效数据集则包含游戏中常用的冲击类声音，如枪声、爆炸音效、拳击声等。
两个数据集均按照90%训练集、10%测试集的比例划分。所有音频均采用44.1kHz采样率、16位量化深度、时长4秒的单声道格式存储。每个数据集的元数据均包含音频索引、所属类别、用于获取Freesound素材链接的ID、音频创作者、音频文件名以及训练/测试集划分信息。
本数据集已用于训练与测试我们在近期投稿且处于审稿阶段的论文中的模型，待论文正式录用后，我们将上传配套的研究论文。