VGMIDI is a dataset of 95 MIDI labelled piano pieces (966 phrases of 4 bars) and 728 non-labelled ones, all piano arrangements of video game soundtracks. Each piece was annotated by 30 human subjects according to a valence-arousal (dimensional) model of emotion. The sentiment of each piece was then extracted by summarizing the 30 annotations and mapping the valence axis to sentiment.
We designed a custom web tool to annotate the video game soundtracks in MIDI format. The source code for this tool can be found here and the details on the annotation process can be found in this paper.
This dataset was presented in this paper, so if you use it, please cite:
@article{ferreira_ismir_2019,
title={Learning to Generate Music with Sentiment},
author={Ferreira, Lucas N. and Whitehead, Jim},
booktitle = {Proceedings of the Conference of the International Society for Music Information Retrieval},
series = {ISMIR'19},
year={2019},
}