MuseScore

在我们的任务中，我们需要一个包含乐器框架标签和多音高标签的大规模数据集。据我们所知，现在不存在这样的数据集。因此，我们从 Musecore 在线社区收集了一个大规模的数据集。该数据集包含各种流派、合成器声音和乐器类别。它还包含每首歌曲的配对 MIDI 和 MP3 文件。我们将 MIDI 文件处理为乐器、音高和钢琴卷标。该数据集的一个主要限制是没有歌声。然而，在未来，我们可以通过将歌声与数据集中的歌曲混合来扩充数据。改进此数据集的另一个计划是从 Musecore 数据集中重新合成 MIDI 文件或添加随机速度以产生更逼真的声音。

In our task, we require a large-scale dataset containing both instrument frame labels and multi-pitch labels. To the best of our knowledge, no such dataset currently exists. Therefore, we collected a large-scale dataset from the Musecore online community. This dataset covers various genres, synthesizer sounds, and instrument categories, and also includes paired MIDI and MP3 files for each song. We processed the MIDI files into instrument, pitch, and piano roll labels. One major limitation of this dataset is the absence of vocal tracks. However, in future work, we can augment the dataset by mixing vocal tracks with the songs within it. Another plan to improve this dataset is to resynthesize MIDI files from the Musecore dataset or add random tempo variations to generate more realistic sounds.

搜集汇总

背景与挑战

背景概述

MuseScore数据集是从Musecore社区收集的大规模音频数据集，涵盖多种流派、合成器声音和乐器类别，提供配对的MIDI和MP3文件，并已处理为乐器、音高和钢琴卷标。其主要限制在于缺乏歌声，未来计划通过混合歌声或重新合成MIDI文件来改进数据质量。