Towards CNN-based Acoustic Modeling of Seventh Chords for Automatic Chord Recognition

In this paper, we build upon a recently proposed deep convolutional neural network architecture for automatic chord recognition (ACR).We focus on extending the commonly used major/minor vocabulary (24 classes) to an extended chord vocabulary of seven chord types with a total of 84 classes. In our experiments, we compare joint and separate classification of the chord type and chord root pitch class using one or two separate models, respectively. We perform a large-scale evaluation using various combinations of training and test sets of different timbre complexity. Our results show that ACR with an extended chord vocabulary achieves high f-scores of 0.97 for isolated chord recordings and 0.66 for mixed contemporary popular music recordings. While the joint ACR modeling leads to the best results for isolated instrument recordings, the separate modeling strategy performs best for complex music recordings. Alongside with this paper, we publish a novel dataset for extended-vocabulary chord recognition which consists of synthetically generated isolated recordings of various musical instruments.