Validating Spotify’s ‘Valence’, ‘Energy’, and ‘Danceability’ Audio Features for Music Psychology Research

Music psychology researchers have long been interested in how emotions are expressed in music. Increasingly, researchers employ computationally derived measures of emotion (e.g., valence and arousal) from services such as Spotify. These measures are easy to use, but how they are calculated is not transparent. It is not clear whether these computational measures reflect how people rate perceived emotions in music. Addressing this, our study investigated the validity of Spotify’s automated ratings with subjective ratings from actual listeners. Participants were recruited from first-year university student cohorts and the general public (N = 244). Participants listened to 40 song excerpts, approximately 20-30 seconds in length, and rated tracks on their perception of its mood (valence), energy (arousal), danceability, familiarity, and enjoyment. Participant ratings were compared with Spotify’s automated ratings. We found that Spotify audio features energy and valence were positively associated with human ratings for these factors, with a strong relationship for arousal, and moderate relationship for valence. Spotify danceability, however, was not strongly associated with human ratings of danceability. These findings corroborate the utility of the widely used circumplex model for measuring emotions concisely. We further extend this by validating automated emotion ratings, such as those implemented by Spotify’s system, for use in music psychology, and highlight that if used cautiously, this is a useful tool in research.