Beat and Downbeat Tracking of Symbolic Music Data Using Deep Recurrent Neural Networks

Musical beat tracking is one of the most investigated tasks in music information retrieval (MIR). Research endeavors on this task have mostly been focused on the modeling of audio data representations. In contrast, beat tracking of symbolic music contents (e.g., MIDI, score sheets) has been relatively overlooked in the past years. In this paper, we revisit the task of symbolic music beat tracking and present to solve this task with modern deep learning approaches. To the extent of our knowledge, it is the first time that utilizing deep learning approaches to track beats and downbeats of symbolic music data. The proposed symbolic beat tracking framework performs joint beat and downbeat tracking in a multi-task learning (MTL) manner, and we investigate various types of networks which are based on the recurrent neural networks (RNN), such as bidirectional long short-term memory (BLSTM) network, hierarchical multi-scale (HM) LSTM, and BLSTM with the attention mechanism. In the experiments, a systematic comparison of these networks and state-of-art audio beat tracking methods are performed on the MusicNet dataset. Experiment results show that the BLSTM model trained specifically on symbolic data outperforms the state-of-the-art beat tracking methods utilized on synthesized audio. Such a comparison of performance also indicates the technical challenges in symbolic music beat tracking.