Generating Smooth Mood-Dynamic Playlists with Audio Features and KNN

Users curate music playlists for many purposes, including focus, enjoyment and therapy. Popular music streaming services generate playlists automatically which are constant in genre or mood. We propose a method to automatically create playlists dynamic in both the Arousal-Valence emotion space and the audio features of songs. Our playlist algorithm uses a two-stage approach to sequentially choose songs, employing a K-Nearest Neighbors (KNN) model to gather potential songs based on emotion and analyzing them with acoustic similarity metrics. To evaluate the effectiveness of various audio feature data, KNN parameters, and similarity metrics, we developed a testing protocol which generates playlists that traverse both Arousal-Valence and audio feature spaces. We define evaluation metrics to measure a playlist’s smoothness and evenness using the Pearson correlation coefficient between dimensions and the variance of steps between songs, respectively. Our algorithm successfully creates smooth and evenly-spaced playlists that transition cohesively in both mood and genre. We explore how the choice of audio feature data, similarity metric, and KNN parameters all have an effect on playlists’ smoothness and evenness across these two spaces.