MusicNet

MusicNet 是 330 个免费许可的古典音乐录音的集合，以及超过 100 万个带注释的标签，指示每个录音中每个音符的精确时间、演奏每个音符的乐器以及音符在韵律结构中的位置组成。标签是通过动态时间扭曲从与录音对齐的乐谱中获取的。标签由训练有素的音乐家验证；我们估计标签错误率为 4%。我们向机器学习和音乐社区提供 MusicNet 标签，作为训练模型的资源和比较结果的通用基准。这个数据集是在论文“Learning Features of Music from Scratch”中介绍的。

MusicNet is a collection of 330 freely-licensed classical music recordings, alongside over 1 million annotated labels that specify the exact timing of every note, the instrument used to play each note, and the note's position within the rhythmic structure of each recording. These labels were derived from sheet music aligned to the recordings via dynamic time warping. The labels were validated by trained musicians; we estimate the label error rate to be 4%. We make the MusicNet labels available to the machine learning and music communities as a resource for model training and a standardized benchmark for comparing results. This dataset was introduced in the paper "Learning Features of Music from Scratch".

搜集汇总

背景与挑战

背景概述

MusicNet是一个包含330个古典音乐录音的数据集，带有超过100万个精确标注的音符标签，涵盖时间、乐器和韵律结构信息，标签错误率约为4%。它专为机器学习和音乐研究设计，用于训练模型和作为通用基准，基于2016年华盛顿大学发布的论文。