We propose a deep learning model for joint beat and downbeat estimation. We tackle the task without incorporating a postprocessing network (often dynamic Bayesian networks). By inspecting a state-of-the-art convolutional approach, we propose several reformulations regarding the network architecture and the loss function. For further details, please refer to "Toward Postprocessing-free Neural Networks for Joint Beat and Downbeat Estimation" (ISMIR 2022).
A model pre-trained on the Ballroom dataset (except tracks whose id=0) is provided. The id of each track can be found in splits.
- python >= 3.6.9
- tensorflow >= 2.5.0
- numpy >= 1.19.5
- mir_eval = 0.6