Structured Training for Large-Vocabulary Chord Recognition

Automatic chord recognition systems operating in the large-vocabulary regime must overcome data scarcity: certain classes occur much less frequently than others, and this presents a significant challenge when estimating model parameters. While most systems model the chord recognition task as a (multi-class) classification problem, few attempts have been made to directly exploit the intrinsic structural similarities between chord classes. In this work, we develop a deep convolutional-recurrent model for automatic chord recognition over a vocabulary of 170 classes. To exploit structural relationships between chord classes, the model is trained to produce both the time-varying chord label sequence as well as binary encodings of chord roots and qualities. This binary encoding directly exposes similarities between related classes, allowing the model to learn a more coherent representation of simultaneous pitch content. Evaluations on a corpus of 1217 annotated recordings demonstrate substantial improvements compared to previous models.