Downbeat Tracking Using Beat Synchronous Features with Recurrent Neural Networks

In this paper, we propose a system that extracts the downbeat times from a beat-synchronous audio feature stream of a music piece. Two recurrent neural networks are used as a front-end: the first one models rhythmic content on multiple frequency bands, while the second one models the harmonic content of the signal. The output activations are then combined and fed into a dynamic Bayesian network which acts as a rhythmical language model. We show on seven commonly used datasets of Western music that the system is able to achieve state-of-the-art results.