e-gmd
收藏资源简介:
扩展的Groove MIDI数据集(E-GMD)是一个大规模的对齐MIDI和合成音频鼓表演录音集合,由Google Magenta创建。该数据集基于Groove MIDI数据集(GMD)中的1,150个MIDI表演,通过43种不同的商用鼓组使用VST渲染管道重新渲染,产生了约49,000个音频录音(1,150个表演×43个鼓组),提供了比原始GMD更大的音色多样性,同时保留了其精确的MIDI注释。E-GMD广泛用于自动鼓转录的研究,其中对不同鼓声音的泛化至关重要。数据集包含16kHz单声道WAV文件和标准MIDI文件,以及链接每个音频/MIDI对的CSV元数据文件。数据集分为训练集(约35,000个示例)、验证集(约8,600个示例)和测试集(约6,450个示例),遵循与GMD相同的基于鼓手的划分。E-GMD使用与GMD相同的9个标准鼓乐器类别,映射自Roland TD-11上的22个原始MIDI音高。数据集的主要限制包括所有音频均为VST渲染,鼓组多样性有限,以及16kHz的采样率可能限制高保真应用的使用。
The Extended Groove MIDI Dataset (E-GMD) is a large-scale collection of aligned MIDI and synthesized audio drum performance recordings, created by Google Magenta. Built upon the 1,150 MIDI performances from the original Groove MIDI Dataset (GMD), this dataset was re-rendered using a VST rendering pipeline across 43 distinct commercial drum kits, yielding approximately 49,000 audio recordings (1,150 performances × 43 drum kits). It offers greater timbral diversity than the original GMD while preserving its precise MIDI annotations. E-GMD is widely used in research on automatic drum transcription, where generalization across different drum sounds is critical. The dataset contains 16kHz mono WAV files and standard MIDI files, alongside CSV metadata files that link each audio-MIDI pair. The dataset is split into training set (~35,000 examples), validation set (~8,600 examples), and test set (~6,450 examples), following the same drummer-based partitioning as GMD. E-GMD utilizes the same 9 standard drum instrument categories as GMD, which are mapped from 22 original MIDI pitches on the Roland TD-11. The primary limitations of the dataset include that all audio is VST-rendered, limited drum kit diversity, and a 16kHz sampling rate that may restrict its use in high-fidelity applications.



