Automatic chord recognition based on the probabilistic modeling of diatonic modal harmony

Chords and keys are among the most exhaustive descriptors of songs. In this study we focus on chord and key sequence recognition from an audio signal, in the context of pop and rock music. The system exploits a set of novel probabilistic models that describe the relationship between different aspects of music and their temporal evolution. These models are based on a set of parameters with a musical meaning. The models include two diatonic key modes, Dorian and Mixolydian, besides major and minor modes previously considered in the literature. These four key modes are the most used in western pop and rock music. In order to provide a compact representation of the chord and key sequences, three novel time-varying harmony- based features are here introduced. Given the importance of emotion characterization in music, the three features are here related to the mood perceived in songs. The method outperforms the state-of-the-art in both chord and key recognition tasks. In order to better train our parameters, we create annotations of chords and keys for a new dataset of 62 songs from the first five Robbie Williams’ albums.