MDS corpus
收藏资源简介:
MDS数据集由约翰内斯开普勒大学团队构建,包含Genre、Random和MAEtest三个子集,旨在评估自动音乐转录系统在不同音乐分布偏移下的性能。数据集包含多样化的音乐流派和随机音符序列,通过雅马哈Disklavier三角钢琴录制,总数据量未明确说明但包含严格筛选的MIDI-音频对齐样本。该数据集专为分析声音与音乐维度上的分布偏移而设计,适用于测试AMT系统在非训练分布条件下的泛化能力,尤其关注古典钢琴音乐之外的转录鲁棒性研究。
The MDS dataset was developed by the research team at Johannes Kepler University. It comprises three subsets: Genre, Random, and MAEtest, and is intended to assess the performance of automatic music transcription (AMT) systems under varying music distribution shifts. The dataset encompasses diverse musical genres and random note sequences, all recorded with a Yamaha Disklavier grand piano. While the total size of the dataset remains unspecified, it includes strictly curated MIDI-audio aligned samples. This dataset is specifically designed for analyzing distribution shifts across acoustic and musical dimensions, and serves as a testbed for evaluating the generalization capability of AMT systems under out-of-training-distribution scenarios, with a particular emphasis on research into transcription robustness beyond classical piano music.




