In this project, we construct a symbolic beat tracking system that performs joint beat and downbeat tracking in a multi-task learning (MTL) manner. The proposed models are based on variants of recurrent neural networks (RNN), All the models are implemented with the PyTorch framework.
In the network stage, we consider three types of neural networks: the first is the conventional bidirectional LSTM (BLSTM) network, the second is the Hierarchical Multiscale RNN (HM-RNN), and the third is the BLSTM with attention mechanism. For a more detailed discussion, please check our paper.
- Ubuntu 18.04
- Python 3.6.8
-
Get a copy: Get a copy of this project by cloning the Git repository.
git clone https://github.com/chuang76/symbolic-beat-tracking.git -
Prerequisite: Install all the dependencies from requirements.txt.
pip install -r requirements.txt -
Run:
-
Input: Put the symbolic music input files into
~/symbolic-beat-tracking/input/.
We provide two kinds of input format: MIDI file or csv file. The input features include "start_time", "end_time", and "note" (i.e. pitch). If you choose to utilize csv file as input data, note that the start_time and end_time represent (onset time * sampling rate) and (offset time * sampling rate), respectively. You can acquire symbolic music data in csv format from the MusicNet dataset.start_time end_time note 0 28126 29662 57 1 46557 52702 73 2 60893 109022 76 -
Output: Simply execute the make command, then you can obtain the beat and downbeat tracking results in
~/symbolic-beat-tracking/output/. The tracking results contain the beat and downbeat positions in seconds.
make -
Here is a short example. There are two symbolic music data files 01.csv and 02.mid in the input folder. You can check their beat/downbeat tracking results as follows.
Yi-Chin Chuang, Li Su, "Beat and Downbeat Tracking of Symbolic Music Data Using Deep Recurrent Neural Networks", Proceedings of Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), December 2020.