Multi-Task Learning of Tempo and Beat: Learning One to Improve the Other

In this paper, we propose a multi-task learning approach for simultaneous tempo estimation and beat tracking of musical audio. The system shows state-of-the-art performance for both tasks on a wide range of data, but has another fundamental advantage: due to its multi-task nature, it is not only able to exploit the mutual information of both tasks by learning a common, shared representation, but can also improve one by learning only from the other. The multi-task learning is achieved by globally aggregating the skip connections of a beat tracking system built around temporal convolutional networks, and feeding them into a tempo classification layer. The benefit of this approach is investigated by the inclusion of training data for which tempo-only annotations are available, and which is shown to provide improvements in beat tracking accuracy.