Abstract
In this work, we present the collection, pre-processing, and trial experimentation of a open-sourced Music Emotion Recognition dataset, augmented with profile informa tion of participants. Through Amazon Mechanical Turk, 452 participants labelled their perceived valence and arousal of 54 songs dynamically. Of the 54 songs, 50 were full length songs and 4 were 45 second excerpts labelled by all participants to be used as a benchmark for data filtering. 9 participant profile features were collected, including demographic information, music listening preferences and musical background. Sta tistical significance was found between a few profile groups, however our trial MER experimentation did not benefit from the addition of profile information. Neverthe less, the dataset shows the potential of cultural and musical background in improving perceived emotion prediction