drum-aware4beat

This repo contains the source codes for paper titled “Drum-Aware Ensemble Architecture for Improved Joint Musical Beat and Downbeat Tracking”. | Paper (arXiv) | Github | Pretrained Models |

This paper presents a novel system architecture that integrates blind source separation with joint beat and downbeat tracking in musical audio signals. The source separation module segregates the percussive and non-percussive components of the input signal, over which beat and downbeat tracking are performed separately and then the results are aggregated with a learnable fusion mechanism. This way, the system can adaptively determine how much the tracking result for an input signal should depend on the input’s percussive or non-percussive components. Evaluation on four testing sets that feature different levels of presence of drum sounds shows that the new architecture consistently outperforms the widely-adopted baseline architecture that does not employ source separation.

usage of this repo

inference:

• Pretrained models mentioned in the paper are provided in pretrained_models. You may run Inference.py to see how it works on the assigned test song.
• To test on you own audio, replace the audio_file_path in inference.py with your test song path should work.

training with your own data:

• dataset preparation:
  • Note that this repo only include 10 songs from GTZAN [1] to demonstrate the usage. You may directly run the training scripts (e.g. train_DA1.py) to see how it work. To train your model with more datasets, please see this repo for how to organize the dataset folders and file names.
  • After uploading your datasets to datasets/original/ following the instructions, you may run the follow scripts to get training data prepared:
    • traintest_split.py
    • source_separation_aug4beat.py
    • prepare_dataset.py
• training:
  • Once your dataset is prepared, the dataset information will be saved in train_dataset.ab and valid_dataset.ab in datasets folder.
  • Run script train_bsl.py, train_DA1.py, or train_DA2.py could start the training. You may also modify the experiment parameters within main() of these scripts.
• hmm optimization:
  • As each trained model may have its prefered parameters for HMM post processing, you may follow HMM optimzation section of this repo to find best HMM for your trained model.

Reference

[1] G. Tzanetakis and P. Cook, “Musical genre classification of audio signals,” IEEE Trans. Speech and Audio Processing, vol. 10, no. 5, pp. 293–302, 2002.