Harmony Transformer: Incorporating Chord Segmentation into Harmony Recognition

Musical harmony analysis is usually a process of unfolding and interpreting the hierarchical structure of music. Computational approaches to such structural analysis are still challenging, owing to the fact that the boundary between different harmonic states (such as chord functions) is not explicitly defined in the audio or symbolic music data. It is a novel approach to improve chord recognition by jointly identifying chord change using end-to-end sequence learning. In this paper, we propose the Harmony Transformer, a multi-task music harmony analysis model aiming to improve chord recognition through incorporating chord segmentation into the recognition process. The integration of chord segmentation and chord recognition is implemented with the Transformer, a deep sequential learning model yielding fruitful results in the field of natural language processing. A non-autoregressive decoding framework is also adopted here in aid of concatenating the two highly correlated tasks. Experiments of both chord symbol recognition and functional harmony recognition on audio and symbolic datasets demonstrate that explicitly learning the hierarchical structural information of musical data can facilitate and improve the harmony recognition