Toward postprocessing-free neural networks for joint beat and downbeat estimation

Recent deep learning-based models for estimating beats and downbeats are mainly composed of three successive stages---feature extraction, sequence modeling, and post processing. While such a framework is prevalent in the scenario of sequence labeling tasks and yields promising results in beat and downbeat estimations, it also indicates a shortage of the employed neural networks, given that the post-processing usually provides a notable performance gain over the previous stage. Moreover, the assumption often made for the post-processing is not suitable for many musical pieces. In this work, we attempt to improve the performance of joint beat and downbeat estimation without incorporating the post-processing stage. By inspecting a state-of-the-art approach, we propose reformulations regarding the network architecture and the loss function. We evaluate our model on various music data and show that the proposed methods are capable of improving the baseline approach without the aid of a post-processing stage