Unsupervised analysis of similarities between musicians and musical genres using spectrograms

Abstract

Since the early days of the information era, digital music has been becoming one of the most consumed types of media, introducingthe need for content-based tools that can search, browse, and retrieve music. Here we describe a method that canquantify similarities between musical genres in an unsupervised fashion, and computes networks of similarities between differentmusicians or musical styles. The method works by converting each song to its 2D spectrogram, and then extracting a largeset of 2883 2D numerical content descriptors. The descriptors are weighted by their informativeness, and then the similaritiesbetween the musical styles are measured using the weighted distances between the musical pieces of each pair of musicians orgenres. The similarities between all pairs provide a similarity matrix, which is visualized by a phylogeny. Experiments using23 well known musicians representing seven musical genres show that the algorithm was able to separate the artists into groupsthat are in agreement with their respective musical genres. The analysis was done in an unsupervised fashion, and without anyhuman definition or annotation of the musical styles.