INTERSPEECH 2021 | arXiv | proceedings

For dataset download, environment setup and data preparation, please refer to this repo.

Refer to the following folders for reproducing results in the paper:

[1] Tables 2-4: schluter-cnn

[2] Tables 5,6: leglaive_lstm

[3] Table 7: lstm_scnn_feat

[4] Table 8: enkd_scnn_feat_student-cnn and enkd_scnn_feat_student-lstm

Inside each folder, run main.py for baselines and main_kd.py for knowledge-distillation expts.

See expts.sh for sample runs.

Check results folder to get hyperparameter configs corresponding to highest validation accuracy. The corresponding test metrics are reported in our paper.

If this code was helpful for your research, consider citing:

@inproceedings{paul21b_interspeech,
  author={Soumava Paul and Gurunath Reddy M and K. Sreenivasa Rao and Partha Pratim Das},
  title={{Knowledge Distillation for Singing Voice Detection}},
  year=2021,
  booktitle={Proc. Interspeech 2021},
  pages={4159--4163},
  doi={10.21437/Interspeech.2021-636}
}

We thank Kyungyun Lee for her revisiting-svd repo which proved to be the starting point of our work.