Chord Function Recognition as Latent State Transition

This paper proposes an unsupervised learning of chord classification aiming at an autonomous recognition of chord functions. In this research, we employ hidden semi-Markov model to incorporate music metrical structure, and in addition, we combine the model with neural network components to embed context information such as beat positions and preceding chord sequences. Experimental results show that the added contexts considerably improve the perplexity. With the help of these neural networks, the proposed model automatically learns hidden states that appropriately represent chord categories. To this purpose, we pre-process the dataset minimally; that is, we only transpose pieces so as not to possess key signatures and ignore octave positions in pitch events. We observe the chord categories effectively cover chords that appeared in the corpus. We further show that the transitions between chord categories reflect the difference of tonalities with a tendency consistent with known chord functions.