Collecting annotations for induced musical emotion via online game with a purpose emotify

Abstract

One of the major reasons why music is so enjoyable is its emotional impact. Indexing and searching by emotion would greatly increase the usability of online music collections. However, there is no consensus on the question which model of emotion would fit this task best. Such a model should be easy for listeners to use both to tag and to retrieve emotion, and should lead to unambiguous results. The latter is complicated not only due to linguistic issues, but also because musical emotion is a subjective phenomenon that depends on many extra-musical factors, such as mood of the listener, musical preferences, age, personality. We investigate this problem by creating a game with a purpose Emotify to collect emotional labels for a set of 400 musical excerpts in different genres. We use the Geneva Emotional Music Scales (GEMS) to annotate this corpus. In this technical report we analyze the data produced by the game. We find that the factors that influence induced musical emotion (in the order of decreasing importance) are musical preferences, mood and gender. We measure the agreement of listeners using Cronbach’s alpha and find that it differs hugely per emotional category (amazement, sadness and solemnity are most inconsistent, and tenderness, power and joyful activation - most consistent categories) and does not differ significantly among the four tested musical genres (rock, pop, classical and electronic music).