Joint Beat and Downbeat Tracking with Recurrent Neural Networks

In this paper we present a novel method for jointly extracting beats and downbeats from audio signals. A recurrent neural network operating directly on magnitude spectrograms is used to model the metrical structure of the audio signals at multiple levels and provides an output feature that clearly distinguishes between beats and downbeats. A dynamic Bayesian network is then used to model bars of variable length and align the predicted beat and downbeat positions to the global best solution. We find that the proposed model achieves state-of-the-art performance on a wide range of different musical genres and styles.