TROMPA-MER: an open dataset for personalized music emotion recognition

Abstract We present a platform and dataset to help research on Music Emotion Recognition (MER). developed the Enthusiasts aiming improve gathering analysis of so-called “ground truth” needed as input MER systems. Firstly, our involves engaging participants using citizen science strategies generate music emotion annotations – presents didactic information musical recommendations incentivization, collects data regarding demographics, mood, language from each participant. Participants annotated excerpt with single free-text words (in native language), distinct forced-choice categories, preference, familiarity. Additionally, stated reasons for annotation including those distinctive perception induction. Secondly, was created personalized contains 181 participants, 4721 annotations, 1161 excerpts. To showcase use dataset, we methodology personalization models based active learning. The experiments show evidence that judgment crowd prior knowledge learning allows more effective systems this particular dataset. Our is publicly available invite researchers it testing