Multi-task Learning of Rhythmic and Tonal Properties in Music

The rapid growth of artificial intelligence has enabled its application to a myriad of tasks across varied fields of research. Advances in A.I. approaches to multimedia related tasks have piqued the interest of a wider general public and have entered the mainstream in forms of analytic applications, entertainment software and otherwise. The music domain is not an exception, methods for audio analysis and music generation, using machine learning, specifically deep learning models, are enjoying an increase in popularity due to their success compared to traditional handcrafted signal processing methods.
The goal of this thesis is to investigate the potential of a combined model for two music analysis tasks: beat tracking and chord recognition. The state-of-the-art solutions for both of these problems heavily rely on deep learning, moreover, multi-task approaches for some music analysis tasks have already been proven successful in the past. It is very likely that the two problems overlap in some way and a joint deep learning model could successfully leverage data present for both tasks. Also sharing a model offers practical benefits like reduced training times and pooling of annotation resources which are expensive to produce for both beat tracking and chord recognition.