amaai-lab/merit
收藏资源简介:
MERIT - 因子控制音乐三元组数据集是一个用于音频分类和特征提取任务的音乐数据集,专门设计用于支持MERIT框架,该框架从单个冻结骨干网络中学习三个独立的音乐相似性空间:旋律、节奏和音色。数据集包含由三元组组成的音频文件,每个三元组围绕一个音乐因子(旋律、节奏或音色)构建,包括一个锚点音频文件、五个正样本音频文件和一个负样本音频文件,以及元数据文件。正样本与锚点在目标因子上相似,但在其他音乐维度(如乐器、流派、调性、速度)上随机变化;负样本则在目标因子上与锚点不同。数据集基于MoisesDB的音频内容,并使用JASCO生成部分音频,适用于非商业用途,遵循CC BY-NC-SA 4.0许可证。
MERIT - Factor-Controlled Music Triplet Dataset is a music dataset for audio classification and feature extraction tasks, specifically designed for the MERIT framework, which learns three independent music similarity spaces—melody, rhythm, and timbre—from a single frozen backbone. The dataset consists of audio files organized into triplets, each centered around a musical factor (melody, rhythm, or timbre), including an anchor audio file, five positive audio files, and one negative audio file, along with metadata. Positives share the target factor with the anchor but vary randomly in other musical dimensions (e.g., instrumentation, genre, key, tempo); negatives differ from the anchor on the target factor. The dataset is derived from MoisesDB audio content, with some audio synthesized by JASCO, and is intended for non-commercial use under the CC BY-NC-SA 4.0 license.




